Add files using upload-large-folder tool

parse/train/BkjLkSqxg/BkjLkSqxg.md

@@ -0,0 +1,280 @@
+# LIPNET: END-TO-END SENTENCE-LEVEL LIPREADING
+
+Yannis M. Assael1,†, Brendan Shillingford1,†, Shimon Whiteson1 & Nando de Freitas1,2,3
+
+Department of Computer Science, University of Oxford, Oxford, UK   
+Google DeepMind, London, UK 2   
+CIFAR, Canada 3   
+{yannis.assael,brendan.shillingford, shimon.whiteson,nando.de.freitas}@cs.ox.ac.uk
+
+# ABSTRACT
+
+Lipreading is the task of decoding text from the movement of a speaker’s mouth. Traditional approaches separated the problem into two stages: designing or learning visual features, and prediction. More recent deep lipreading approaches are end-to-end trainable (Wand et al., 2016; Chung & Zisserman, 2016a). However, existing work on models trained end-to-end perform only word classification, rather than sentence-level sequence prediction. Studies have shown that human lipreading performance increases for longer words (Easton & Basala, 1982), indicating the importance of features capturing temporal context in an ambiguous communication channel. Motivated by this observation, we present LipNet, a model that maps a variable-length sequence of video frames to text, making use of spatiotemporal convolutions, a recurrent network, and the connectionist temporal classification loss, trained entirely end-to-end. To the best of our knowledge, LipNet is the first end-to-end sentence-level lipreading model that simultaneously learns spatiotemporal visual features and a sequence model. On the GRID corpus, LipNet achieves $9 5 . 2 \%$ accuracy in sentence-level, overlapped speaker split task, outperforming experienced human lipreaders and the previous $8 6 . 4 \%$ word-level state-of-the-art accuracy (Gergen et al., 2016).
+
+# 1 INTRODUCTION
+
+Lipreading plays a crucial role in human communication and speech understanding, as highlighted by the McGurk effect (McGurk & MacDonald, 1976), where one phoneme’s audio dubbed on top of a video of someone speaking a different phoneme results in a third phoneme being perceived.
+
+Lipreading is a notoriously difficult task for humans, specially in the absence of context1. Most lipreading actuations, besides the lips and sometimes tongue and teeth, are latent and difficult to disambiguate without context (Fisher, 1968; Woodward & Barber, 1960). For example, Fisher (1968) gives 5 categories of visual phonemes (called visemes), out of a list of 23 initial consonant phonemes, that are commonly confused by people when viewing a speaker’s mouth. Many of these were asymmetrically confused, and observations were similar for final consonant phonemes.
+
+Consequently, human lipreading performance is poor. Hearing-impaired people achieve an accuracy of only $1 7 \pm \mathrm { \dot { 1 } 2 \% }$ even for a limited subset of 30 monosyllabic words and $2 1 \pm 1 1 \%$ for 30 compound words (Easton & Basala, 1982). An important goal, therefore, is to automate lipreading. Machine lipreaders have enormous practical potential, with applications in improved hearing aids, silent dictation in public spaces, security, speech recognition in noisy environments, biometric identification, and silent-movie processing.
+
+Machine lipreading is difficult because it requires extracting spatiotemporal features from the video (since both position and motion are important). Recent deep learning approaches attempt to extract those features end-to-end. Most existing work, however, performs only word classification, not sentence-level sequence prediction.
+
+In this paper, we present LipNet, which is to the best of our knowledge, the first end-to-end sentence-level lipreading model. As with modern deep learning based automatic speech recognition (ASR), LipNet is trained end-to-end to make sentence-level predictions. Our model operates at the character-level, using spatiotemporal convolutional neural networks (STCNNs), recurrent neural networks (RNNs), and the connectionist temporal classification loss (CTC) Graves et al. (2006).
+
+Our empirical results on the GRID corpus (Cooke et al., 2006), one of the few public sentence-level datasets, show that LipNet attains a $9 \bar { 5 } . 2 \%$ sentence-level word accuracy, in a overlapped speakers split that is popular for benchmarking lipreading methods. The previous best accuracy reported on an aligned word classification version of this task was $8 6 . 4 \%$ (Gergen et al., 2016). Furthermore, LipNet can generalise across unseen speakers in the GRID corpus with an accuracy of $8 8 . 6 \%$ .
+
+We also compare the performance of LipNet with that of hearing-impaired people who can lipread on the GRID corpus task. On average, they achieve an accuracy of ${ \bar { 5 } } 2 . 3 \%$ , in contrast to LipNet’s $1 . 6 9 \times$ higher accuracy in the same sentences.
+
+Finally, by applying saliency visualisation techniques (Zeiler & Fergus, 2014; Simonyan et al., 2013), we interpret LipNet’s learned behaviour, showing that the model attends to phonologically important regions in the video. Furthermore, by computing intra-viseme and inter-viseme confusion matrices at the phoneme level, we show that almost all of LipNet’s few erroneous predictions occur within visemes, since context is sometimes insufficient for disambiguation.
+
+# 2 RELATED WORK
+
+In this section, we outline various existing approaches to automated lipreading.
+
+Automated lipreading: Most existing work on lipreading does not employ deep learning. Such work requires either heavy preprocessing of frames to extract image features, temporal preprocessing of frames to extract video features (e.g., optical flow or movement detection), or other types of handcrafted vision pipelines (Matthews et al., 2002; Zhao et al., 2009; Gurban & Thiran, 2009; Papandreou et al., 2007; 2009; Pitsikalis et al., 2006; Lucey & Sridharan, 2006; Papandreou et al., 2009). The automated lipreading literature is too vast to adequately cover, so we refer the reader to Zhou et al. (2014) for an extensive review.
+
+Notably, Goldschen et al. (1997) were the first to do visual-only sentence-level lipreading using hidden Markov models (HMMs) in a limited dataset, using hand-segmented phones. Later, Neti et al. (2000) were the first to do sentence-level audiovisual speech recognition using an HMM combined with hand-engineered features, on the IBM ViaVoice (Neti et al., 2000) dataset. The authors improve speech recognition performance in noisy environments by fusing visual features with audio ones. The dataset contains 17111 utterances of 261 speakers for training (about 34.9 hours) and is not publicly available. As stated, their visual-only results cannot be interpreted as visual-only recognition, as they are used as rescoring of the noisy audio-only lattices. Using a similar approach, Potamianos et al. (2003) report speaker independent and speaker adapted $9 1 . 6 \hat { 2 } \%$ , $8 2 . 3 1 \%$ WER in the same dataset respectively, and $3 8 . 5 3 \%$ , $1 6 . 7 7 \%$ WER in the connected DIGIT corpus, which contains sentences of digits.
+
+Furthermore, Gergen et al. (2016) use speaker-dependent training on an LDA-transformed version of the Discrete Cosine Transforms of the mouth regions in an HMM/GMM system. This work holds the previous state-of-the-art on the GRID corpus with a speaker-dependent accuracy of $8 6 . 4 \%$ . Generalisation across speakers and extraction of motion features is considered an open problem, as noted in (Zhou et al., 2014). LipNet addresses both of these issues.
+
+Classification with deep learning: In recent years, there have been several attempts to apply deep learning to lipreading. However, all of these approaches perform only word or phoneme classification, whereas LipNet performs full sentence sequence prediction. Approaches include learning multimodal audio-visual representations (Ngiam et al., 2011; Sui et al., 2015; Ninomiya et al., 2015; Petridis & Pantic, 2016), learning visual features as part of a traditional speech-style processing pipeline (e.g. HMMs, GMM-HMMs, etc.) for classifying words and/or phonemes (Almajai et al., 2016; Takashima et al., 2016; Noda et al., 2014; Koller et al., 2015), or combinations thereof (Takashima et al., 2016). Many of these approaches mirror early progress in applying neural networks for acoustic processing in speech recognition (Hinton et al., 2012).
+
+Chung & Zisserman (2016a) propose spatial and spatiotemporal convolutional neural networks, based on VGG, for word classification. The architectures are evaluated on a word-level dataset BBC TV (333 and 500 classes), but, as reported, their spatiotemporal models fall short of the spatial architectures by an average of around $1 4 \%$ . Additionally, their models cannot handle variable sequence lengths and they do not attempt sentence-level sequence prediction.
+
+Chung & Zisserman (2016b) train an audio-visual max-margin matching model for learning pretrained mouth features, which they use as inputs to an LSTM for 10-phrase classification on the OuluVS2 dataset, as well as a non-lipreading task.
+
+Wand et al. (2016) introduce LSTM recurrent neural networks for lipreading but address neither sentence-level sequence prediction nor speaker independence.
+
+Garg et al. (2016) apply a VGG pre-trained on faces to classifying words and phrases from the MIRACL-VC1 dataset, which has only 10 words and 10 phrases. However, their best recurrent model is trained by freezing the VGGNet parameters and then training the RNN, rather than training them jointly. Their best model achieves only $5 6 . 0 \%$ word classification accuracy, and $4 4 . 5 \%$ phrase classification accuracy, despite both of these being 10-class classification tasks.
+
+Sequence prediction in speech recognition: The field of automatic speech recognition (ASR) would not be in the state it is today without modern advances in deep learning, many of which have occurred in the context of ASR (Graves et al., 2006; Dahl et al., 2012; Hinton et al., 2012). The connectionist temporal classification loss (CTC) of Graves et al. (2006) drove the movement from deep learning as a component of ASR, to deep ASR systems trained end-to-end (Graves & Jaitly, 2014; Maas et al., 2015; Amodei et al., 2015). As mentioned earlier, much recent lipreading progress has mirrored early progress in ASR, but stopping short of sequence prediction.
+
+LipNet is the first end-to-end model that performs sentence-level sequence prediction for visual speech recogntion. That is, we demonstrate the first work that takes as input as sequence of images and outputs a distribution over sequences of tokens; it is trained end-to-end using CTC and thus also does not require alignments.
+
+Lipreading Datasets: Lipreading datasets (AVICar, AVLetters, AVLetters2, BBC TV, CUAVE, OuluVS1, OuluVS2) are plentiful (Zhou et al., 2014; Chung & Zisserman, 2016a), but most only contain single words or are too small. One exception is the GRID corpus (Cooke et al., 2006), which has audio and video recordings of 34 speakers who produced 1000 sentences each, for a total of 28 hours across 34000 sentences. Table 1 summarises state-of-the-art performance in each of the main lipreading datasets.
+
+Table 1: Existing lipreading datasets and the state-of-the-art accuracy reported on these. The size column represents the number of utterances used by the authors for training. Although the GRID corpus contains entire sentences, Gergen et al. (2016) consider only the simpler case of predicting isolated words. LipNet predicts sequences and hence can exploit temporal context to attain much higher accuracy. Phrase-level approaches were treated as plain classification.   
+
+<table><tr><td>Method</td><td>Dataset</td><td>Size</td><td>Output</td><td>Accuracy</td></tr><tr><td colspan="5"></td></tr><tr><td>Fu et al. (2008)</td><td>AVICAR</td><td>851</td><td>Digits</td><td>37.9%</td></tr><tr><td>Hu et al. (2016)</td><td>AVLetter</td><td>78</td><td>Alphabet</td><td>64.6%</td></tr><tr><td>Papandreou et al. (2009)</td><td>CUAVE</td><td>1800</td><td>Digits</td><td>83.0%</td></tr><tr><td>Chung &amp; Zisserman (2016a)</td><td>OuluVS1</td><td>200</td><td>Phrases</td><td>91.4%</td></tr><tr><td>Chung &amp; Zisserman (2016b)</td><td>OuluVS2</td><td>520</td><td>Phrases</td><td>94.1%</td></tr><tr><td>Chung &amp; Zisserman ( 1 (2016a)</td><td>BBC TV</td><td>&gt; 400000</td><td>Words</td><td>65.4%</td></tr><tr><td>Gergen et al. (2016)</td><td>GRID</td><td>29700</td><td>Words*</td><td>86.4%</td></tr><tr><td>LipNet</td><td>GRID</td><td>28775</td><td>Sentences</td><td>95.2%</td></tr></table>
+
+We use the GRID corpus to evaluate LipNet because it is sentence-level and has the most data. The sentences are drawn from the following simple grammar: command $^ { ( 4 ) } + c o l o r ^ { ( 4 ) } +$ prepositio $\ l _ { \ l } ^ { ( 4 ) } + l e t t e r ^ { ( 2 5 ) } + d i g i t ^ { ( 1 0 ) } + a d v e r ^ { \ l } \ l _ { b } ^ { \ l ^ { ( 4 ) } }$ , where the number denotes how many word choices there are for each of the 6 word categories. The categories consist of, respectively, {bin, lay, place, set}, {blue, green, red, white}, {at, by, in, with}, $\{ \bar { A } , \ldots , Z \} \backslash \{ W \}$ , $\{ \mathrm { z e r o } , \ldots , \mathrm { n i n e } \}$ , and {again, now, please, $\operatorname { s o o n } \}$ , yielding 64000 possible sentences. For example, two sentences in the data are “set blue by A four please” and “place red at C zero again”.
+
+# 3 LIPNET
+
+LipNet is a neural network architecture for lipreading that maps variable-length sequences of video frames to text sequences, and is trained end-to-end. In this section, we describe LipNet’s building blocks and architecture.
+
+# 3.1 SPATIOTEMPORAL CONVOLUTIONS
+
+Convolutional neural networks (CNNs), containing stacked convolutions operating spatially over an image, have been instrumental in advancing performance in computer visions tasks such as object recognition that receive an image as input (Krizhevsky et al., 2012). A basic 2D convolution layer from $C$ channels to $C ^ { \prime }$ channels (without a bias and with unit stride) computes
+
+$$
+[ \mathrm { c o n v } ( { \bf x } , { \bf w } ) ] _ { c ^ { \prime } i j } = \sum _ { c = 1 } ^ { C } \sum _ { i ^ { \prime } = 1 } ^ { k _ { w } } \sum _ { j ^ { \prime } = 1 } ^ { k _ { h } } w _ { c ^ { \prime } c i ^ { \prime } j ^ { \prime } } x _ { c , i + i ^ { \prime } , j + j ^ { \prime } } ,
+$$
+
+for input $\mathbf { x }$ and weights $\mathbf { w } \in \mathbb { R } ^ { C ^ { \prime } \times C \times k _ { w } \times k _ { h } }$ where we define $x _ { c i j } = 0$ for $i , j$ out of bounds. Spatiotemporal convolutional neural networks (STCNNs) can process video data by convolving across time, as well as the spatial dimensions (Karpathy et al., 2014; Ji et al., 2013). Hence similarly,
+
+$$
+[ \mathrm { s t c o n v } ( { \bf x } , { \bf w } ) ] _ { c ^ { \prime } t i j } = \sum _ { c = 1 } ^ { C } \sum _ { t ^ { \prime } = 1 } ^ { k _ { t } } \sum _ { i ^ { \prime } = 1 } ^ { k _ { w } } \sum _ { j ^ { \prime } = 1 } ^ { k _ { h } } w _ { c ^ { \prime } c t ^ { \prime } i ^ { \prime } j ^ { \prime } } x _ { c , t + t ^ { \prime } , i + i ^ { \prime } , j + j ^ { \prime } } .
+$$
+
+# 3.2 GATED RECURRENT UNIT
+
+Gated Recurrent Unit (GRU) (Chung et al., 2014) is a type of recurrent neural network (RNN) that improves upon earlier RNNs by adding cells and gates for propagating information over more timesteps and learning to control this information flow. It is similar to the Long Short-Term Memory (LSTM) RNN (Hochreiter & Schmidhuber, 1997). We use the standard formulation:
+
+$$
+\begin{array} { r l } & { [ \mathbf { u } _ { t } , \mathbf { r } _ { t } ] ^ { T } = \mathrm { s i g m } ( \mathbf { W } _ { z } \mathbf { z } _ { t } + \mathbf { W } _ { h } \mathbf { h } _ { t - 1 } + \mathbf { b } _ { g } ) } \\ & { \qquad \tilde { \mathbf { h } } _ { t } = \mathrm { t a n h } ( \mathbf { U } _ { z } \mathbf { z } _ { t } + \mathbf { U } _ { h } ( \mathbf { r } _ { t } \odot \mathbf { h } _ { t - 1 } ) + \mathbf { b } _ { h } ) } \\ & { \qquad \mathbf { h } _ { t } = ( \mathbf { 1 } - \mathbf { u } _ { t } ) \odot \mathbf { h } _ { t - 1 } + \mathbf { u } _ { t } \odot \tilde { \mathbf { h } } _ { t } } \end{array}
+$$
+
+where $\mathbf { z } : = \{ \mathbf { z } _ { 1 } , \dots , \mathbf { z } _ { T } \}$ is the input sequence to the RNN, $\odot$ denotes element-wise multiplication, and $\mathrm { s i g m } ( r ) = 1 / ( 1 + \exp ( - r ) )$ . We use a bidirectional GRU (Bi-GRU) as introduced by Graves & Schmidhuber (2005) in the context of LSTMs: one RNN maps $\{ \mathbf { z } _ { 1 } , \dotsc , \mathbf { z } _ { T } \} \mapsto \{ \overrightarrow { \mathbf { h } _ { 1 } } , \dotsc , \overrightarrow { \mathbf { h } _ { T } } \}$ , and another $\{ \mathbf { z } _ { T } , \dotsc , \mathbf { z } _ { 1 } \} \mapsto \{ { \overleftarrow { \mathbf { h } _ { 1 } } } , \dotsc , { \overleftarrow { \mathbf { h } _ { T } } } \}$ , then $\mathbf { h } _ { t } : = [ \overrightarrow { \mathbf { h } _ { t } } , \overleftarrow { \mathbf { h } _ { t } } ]$ . The Bi-GRU ensures that $\mathbf { h } _ { t }$ depends on $\mathbf { z } _ { t ^ { \prime } }$ for all $t ^ { \prime }$ . To parameterise a distribution over sequences, at time-step $t$ let $p ( u _ { t } | \mathbf { z } ) =$ softmax $( \mathrm { m l p } ( \mathbf { h } _ { t } ; \mathbf { W } _ { m l p } ) )$ , where mlp is a feed-forward network with weights $\mathbf { W } _ { m l p }$ . Then we can define the distribution over length- $\mathcal { T }$ sequences as $\begin{array} { r } { p ( u _ { 1 } , \ldots , u _ { T } | \mathbf { z } ) = \prod _ { 1 \leq t \leq T } p ( u _ { t } | \mathbf { \bar { z } } ) } \end{array}$ , where $T$ is determined by $\mathbf { z }$ , the input to the GRU. In LipNet, $\mathbf { z }$ is the output of the STCNN.
+
+# 3.3 CONNECTIONIST TEMPORAL CLASSIFICATION
+
+The connectionist temporal classification (CTC) loss (Graves et al., 2006) is widely used in modern speech recognition as it eliminates the need for training data that aligns inputs to target outputs (Amodei et al., 2015; Graves & Jaitly, 2014; Maas et al., 2015). Given a model that outputs a sequence of discrete distributions over the token classes (vocabulary) augmented with a special “blank” token, CTC computes the probability of a sequence by marginalising over all sequences that are defined as equivalent to this sequence. This simultaneously removes the need for alignments and addresses variable-length sequences. Let $V$ denote the set of tokens that the model classifies at a single time-step of its output (vocabulary), and the blank-augmented vocabulary ${ \tilde { V } } = V \cup \{ \ldots \}$ where denotes the CTC blank symbol. Define the function $B : \tilde { V } ^ { * } \to V ^ { * }$ that, given a string over $\tilde { V }$ , deletes adjacent duplicate characters and removes blank tokens. For a label sequence $y \in V ^ { * }$ , CTC defines $\begin{array} { r } { \dot { p } ( y | \mathbf { x } ) = \sum _ { u \in \mathcal { B } ^ { - 1 } ( y ) } } \end{array}$ s.t. $_ { | u | = T } p \big ( u _ { 1 } , \dots , u _ { T } | \mathbf x \big )$ , where $T$ is the number of time-steps in the sequence model. For example, if $T = 3$ , CTC defines the probability of a string “am” as $p ( a a m ) \bar { + } p ( a m m ) + p ( . a m ) + \bar { p } ( a . m ) + p ( a m . )$ . This sum is computed efficiently by dynamic programming, allowing us to perform maximum likelihood.
+
+![](images/1872721e1db4aa3d870c4e29c3ac890f371742e27758f736f3657583d4e43b7e.jpg)  
+Figure 1: LipNet architecture. A sequence of $T$ frames is used as input, and is processed by 3 layers of STCNN, each followed by a spatial max-pooling layer. The features extracted are processed by 2 Bi-GRUs; each time-step of the GRU output is processed by a linear layer and a softmax. This end-to-end model is trained with CTC.
+
+# 3.4 LIPNET ARCHITECTURE
+
+Figure 1 illustrates the LipNet architecture, which starts with $3 \times$ (spatiotemporal convolutions, channel-wise dropout, spatial max-pooling). Subsequently, the features extracted are followed by two Bi-GRUs. The Bi-GRUs are crucial for efficient further aggregation of the STCNN output. Finally, a linear transformation is applied at each time-step, followed by a softmax over the vocabulary augmented with the CTC blank, and then the CTC loss. All layers use rectified linear unit (ReLU) activation functions. More details including hyperparameters can be found in Table 3 of Appendix A.
+
+# 4 LIPREADING EVALUATION
+
+In this section, we evaluate LipNet on the GRID corpus. The augmentation methods employed don’t make use of external data and rely purely on the GRID corpus.
+
+# 4.1 DATA AUGMENTATION
+
+Preprocessing: The GRID corpus consists of 34 subjects, each narrating 1000 sentences. The videos for speaker 21 are missing, and a few others are empty or corrupt, leaving 32746 usable videos. We employ a split (unseen speakers; not previously used in the literature) holding out the data of two male speakers (1 and 2) and two female speakers (20 and 22) for evaluation (3971 videos). The remainder is used for training (28775 videos). We also use a sentence-level variant of the split (overlapped speakers) similar to Wand et al. (2016), where 255 random sentences from each speaker are used for evaluation. All remaining data from all speakers is pooled together for training. All videos are 3 seconds long with a frame rate of 25fps. The videos were processed with the DLib face detector, and the iBug face landmark predictor (Sagonas et al., 2013) with 68 landmarks coupled with an online Kalman Filter. Using these landmarks, we apply an affine transformation to extract a mouth-centred crop of size $1 0 0 \times 5 0$ pixels per frame. We standardise the RGB channels over the whole training set to have zero mean and unit variance.
+
+Augmentation: We augment the dataset with simple transformations to reduce overfitting. First, we train on both the regular and the horizontally mirrored image sequence. Second, since the dataset provides word start and end timings for each sentence video, we augment the sentence-level training data with video clips of individual words as additional training instances. These instances have a decay rate of 0.925. Third, to encourage resilience to varying motion speeds by deletion and duplication of frames, this is performed with a per-frame probability of 0.05. The same augmentation methods were followed in all proposed baselines and models.
+
+# 4.2 BASELINES
+
+To evaluate LipNet, we compare its performance to that of three hearing-impaired people who can lipread, as well as three ablation models inspired by recent state-of-the-art work (Chung & Zisserman, 2016a; Wand et al., 2016).
+
+Hearing-Impaired People: This baseline was performed by three members of the Oxford Students’ Disability Community. After being introduced to the grammar of the GRID corpus, they observed 10 minutes of annotated videos from the training dataset, then annotated 300 random videos from the evaluation dataset. When uncertain, they were asked to pick the most probable answer.
+
+Baseline-LSTM: Using the sentence-level training setup of LipNet, we replicate the model architecture of the previous deep learning GRID corpus state-of-the-art (Wand et al., 2016). See Appendix A for more implementation details.
+
+Baseline-2D: Based on the LipNet architecture, we replace the STCNN with spatial-only convolutions similar to those of Chung & Zisserman (2016a). Notably, contrary to the results we observe with LipNet, Chung & Zisserman (2016a) report $1 4 \%$ and $3 1 \%$ poorer performance of their STCNNs compared to the 2D architectures in their two datasets.
+
+Baseline-NoLM: Identical to LipNet, but with the language model used in beam search disabled.
+
+# 4.3 PERFORMANCE EVALUATION
+
+To measure the performance of LipNet and the baselines, we compute the word error rate (WER) and the character error rate (CER), standard metrics for the performance of ASR models. We produce approximate maximum-probability predictions from LipNet by performing CTC beam search. WER (or CER) is defined as the minimum number of word (or character) insertions, substitutions, and deletions required to transform the prediction into the ground truth, divided by the number of words (or characters) in the ground truth. Note that WER is usually equal to classification error when the predicted sentence has the same number of words as the ground truth, particularly in our case since almost all errors are substitution errors.
+
+Table 2 summarises the performance of LipNet compared to the baselines. According to the literature, the accuracy of human lipreaders is around $2 0 \%$ (Easton & Basala, 1982; Hilder et al., 2009). As expected, the fixed sentence structure and the limited subset of words for each position in the GRID corpus facilitate the use of context, increasing performance. On the unseen speakers split, the three hearing-impaired people achieve $5 7 . 3 \%$ , $5 0 . { \overset { - } { 4 } } { \overset { - } { \% } }$ , and $3 5 . 5 \%$ WER respectively, yielding an average of $4 7 . 7 \%$ WER.
+
+Table 2: Performance of LipNet on the GRID dataset compared to the baselines, measured on two splits: (a) evaluating on only unseen speakers, and (b) evaluating on a 255 video subset of each speakers’ sentences.   
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Unseen Speakers</td><td colspan="2">Overlapped Speakers</td></tr><tr><td>CER</td><td>WER</td><td>CER</td><td>WER</td></tr><tr><td>Hearing-Impaired Person (avg)</td><td></td><td>47.7%</td><td></td><td></td></tr><tr><td>Baseline-LSTM</td><td>38.4%</td><td>52.8%</td><td>15.2%</td><td>26.3%</td></tr><tr><td>Baseline-2D</td><td>16.2%</td><td>26.7%</td><td>4.3%</td><td>11.6%</td></tr><tr><td>Baseline-NoLM</td><td>6.7%</td><td>13.6%</td><td>2.0%</td><td>5.6%</td></tr><tr><td>LipNet</td><td>6.4%</td><td>11.4%</td><td>1.9%</td><td>4.8%</td></tr></table>
+
+For both unseen and overlapped speakers evaluation, the highest performance is achieved by the architectures enhanced with convolutional stacks. LipNet exhibits a $2 . 3 \times$ higher performance in the overlapped compared to the unseen speakers split. For unseen speakers, Baseline-2D and LipNet achieve $1 . 8 \times$ and $4 . 2 \times$ lower WER, respectively, than hearing-impaired people.
+
+The WER for unseen speakers Baseline-2D is $2 6 . 7 \%$ , whereas for LipNet it is $2 . 3 \times$ lower, at $1 1 . 4 \%$ . Similarly, the error rate for overlapped speakers was $2 . 4 \times$ lower for LipNet compared to Baseline2D. Both results demonstrate the importance of combining STCNNs with RNNs. This performance difference confirms the intuition that extracting spatiotemporal features using a STCNN is better than aggregating spatial-only features. This observation contrasts with the empirical observations of Chung & Zisserman (2016a). Furthermore, LipNet’s use of STCNN, RNNs, and CTC cleanly allow processing both variable-length input and variable-length output sequences, whereas the architectures of Chung & Zisserman (2016a) and Chung & Zisserman (2016b) only handle the former.
+
+Baseline-LSTM exhibits the lowest performance, in both unseen and overlapped speakers, with $5 2 . 8 \%$ and $2 6 . 3 \%$ WER, respectively. Interestingly, although Baseline-LSTM replicates the architecture of Wand et al. (2016), and despite the numerous data augmentation methods, the model performs $1 . 3 \times$ lower than the reported $7 9 . 6 \%$ word-level accuracy illustrating the difficulty of a sentence-level task even in a restricted grammar.
+
+Finally, by disabling the language model, the Baseline-NoLM exhibits approximately $1 . 2 \times$ higher WER than our proposed model.
+
+# 4.4 LEARNED REPRESENTATIONS
+
+In this section, we analyse the learned representations of LipNet from a phonological perspective. First, we create saliency visualisations (Simonyan et al., 2013; Zeiler & Fergus, 2014) to illustrate where LipNet has learned to attend. In particular, we feed an input into the model and greedily decode an output sequence, yielding a CTC alignment $\hat { u } \in \tilde { V } ^ { * }$ (following the notation of Sections 3.2 and 3.3). Then, we compute the gradient of $\textstyle \sum _ { t } p ( { \hat { u } } _ { t } | \mathbf { x } )$ with respect to the input video frame sequence, but unlike Simonyan et al. (2013), we use guided backpropagation (Springenberg et al., 2014). Second, we train LipNet to predict ARPAbet phonemes, instead of characters, to analyse visual phoneme similarities using intra-viseme and inter-viseme confusion matrices.
+
+# 4.4.1 SALIENCY MAPS
+
+We apply saliency visualisation techniques to interpret LipNet’s learned behaviour, showing that the model attends to phonologically important regions in the video. In particular, in Figure 2 we analyse two saliency visualisations for the words please and lay for speaker 25, based on Ashby (2013).
+
+![](images/d6845168838f936d340381ba209455b7c18df7f8eeafe691226d8b5dedc3de4b.jpg)  
+Figure 2: Saliency maps for the words (a) please and (b) lay, produced by backpropagation to the input, showing the places where LipNet has learned to attend. The pictured transcription is given by greedy CTC decoding. CTC blanks are denoted by $\cdot \_$ .
+
+The production of the word please requires a great deal of articulatory movement at the beginning: the lips are pressed firmly together for the bilabial plosive $/ \mathrm { p } /$ (frame 1). At the same time, the blade of the tongue comes in contact with the alveolar ridge in anticipation of the following lateral $/ \mathrm { { l } } / .$ . The lips then part, allowing the compressed air to escape between the lips (frame 2). The jaw and lips then open further, seen in the distance between the midpoints of the upper and lower lips, and the lips spread (increasing the distance between the corners of the mouth), for the close vowel /iy/ (frame 3–4). Since this is a relatively steady-state vowel, lip position remains unchanged for the rest of its duration (frames 4–8), where the attention level drops considerably. The jaw and the lips then close slightly, as the blade of the tongue needs to be brought close to the alveolar ridge, for $/ \mathbf { Z } /$ (frames 9–10), where attention resumes.
+
+Lay is interesting since the bulk of frontally visible articulatory movement involves the blade of the tongue coming into contact with the alveolar ridge for $/ \mathrm { l } / $ (frames 2–6), and then going down for the vowel /ey/ (frames 7–9). That is exactly where most of LipNet’s attention is focused, as there is little change in lip position.
+
+# 4.4.2 VISEMES
+
+According to DeLand (1931) and Fisher (1968), Alexander Graham Bell first hypothesised that multiple phonemes may be visually identical on a given speaker. This was later verified, giving rise to the concept of a viseme, a visual equivalent of a phoneme (Woodward & Barber, 1960; Fisher, 1968). For our analysis, we use the phoneme-to-viseme mapping of Neti et al. (2000), clustering the phonemes into the following categories: Lip-rounding based vowels (V), Alveolar-semivowels (A), Alveolar-fricatives (B), Alveolar (C), Palato-alveolar (D), Bilabial (E), Dental (F), Labio-dental (G), and Velar (H). The full mapping can be found in Table 4 in Appendix A. The GRID corpus contain 31 out of the 39 phonemes in ARPAbet. We compute confusion matrices between phonemes and then group phonemes into viseme clusters, following Neti et al. (2000). Figure 3 shows the confusion matrices of the 3 most confused viseme categories, as well as the confusions between the viseme categories. The full phoneme confusion matrix is in Figure 4 in Appendix B.
+
+![](images/3feedcc8be2ef8745cfe831ae3909203f4444d251cb9705a44e0d9f421116203.jpg)  
+Figure 3: Intra-viseme and inter-viseme confusion matrices, depicting the three categories with the most confusions, as well as the confusions between viseme clusters. Colours are row-normalised to emphasise the errors.
+
+Given that the speakers are British, the confusion between /aa/ and /ay/ (Figure 3a) is most probably due to the fact that the first element, and the greater part, of the diphthong /ay/ is articulatorily identical with /aa/: an open back unrounded vowel (Ferragne & Pellegrino, 2010). The confusion of /ih/ (a rather close vowel) and /ae/ (a very open vowel) is at first glance surprising, but in fact in the sample /ae/ occurs only in the word at, which is a function word normally pronounced with a reduced, weak vowel /ah/. /ah/ and $/ \mathrm { i h } I$ are the most frequent unstressed vowels and there is a good deal of variation within and between them, e.g. private and watches (Cruttenden, 2014).
+
+The confusion within the categories of bilabial stops $/ { \mathrm { p } } { \mathrm { ~ b ~ m } } ^ { \prime }$ and alveolar stops /t d n/ (Figures 3b-c) is unsurprising: complete closure at the same place of articulation makes them look practically identical. The differences of velum action and vocal fold vibration are unobservable from the front.
+
+Finally, the quality of the viseme categorisation of Neti et al. (2000) is confirmed by the fact that the matrix in Figure 3d is diagonal, with only minor confusion between alveolar (C) and palatoalveolar (D) visemes. Articulatorily, alveolar $/ s \ z /$ and palato-alveolar $/ \mathrm { s h \ z h } /$ fricatives are distinguished by only a small difference in tongue position: against the palate just behind the alveolar ridge, which is not easily observed from the front. The same can be said about dental /th/ and alveolar $/ \mathrm { t } / $ .
+
+# 5 CONCLUSIONS
+
+We proposed LipNet, the first model to apply deep learning to end-to-end learning of a model that maps sequences of image frames of a speaker’s mouth to entire sentences. The end-to-end model eliminates the need to segment videos into words before predicting a sentence. LipNet requires neither hand-engineered spatiotemporal visual features nor a separately-trained sequence model.
+
+Our empirical evaluation illustrates the importance of spatiotemporal feature extraction and efficient temporal aggregation, confirming the intuition of Easton & Basala (1982). Furthermore, LipNet greatly outperforms a human lipreading baseline, exhibiting $4 . 1 \times$ better performance, and $\bar { 4 } . 8 \%$ WER which is $2 . 8 \times$ lower than the word-level state-of-the-art (Gergen et al., 2016) in the GRID corpus.
+
+While LipNet is already an empirical success, the deep speech recognition literature (Amodei et al., 2015) suggests that performance will only improve with more data. In future work, we hope to demonstrate this by applying LipNet to larger datasets, such as a sentence-level variant of that collected by Chung & Zisserman (2016a).
+
+Some applications, such as silent dictation, demand the use of video only. However, to extend the range of potential applications of LipNet, we aim to apply this approach to a jointly trained audiovisual speech recognition model, where visual input assists with robustness in noisy environments.
+
+# ACKNOWLEDGMENTS
+
+This work was supported by an Oxford-Google DeepMind Graduate Scholarship, the EPSRC, and CIFAR. We would also like to thank: NVIDIA for their generous donation of DGX-1 and GTX Titan X GPUs, used in our experiments; Aine Jackson, Brittany Klug and Samantha Pugh for helping ´ us measure the experienced lipreader baseline; Mitko Sabev for his phonetics guidance; Odysseas Votsis for his video production help; and Alex Graves and Oiwi Parker Jones for helpful comments.
+
+# REFERENCES
+
+I. Almajai, S. Cox, R. Harvey, and Y. Lan. Improved speaker independent lip reading using speaker adaptive training and deep neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 2722–2726, 2016.   
+D. Amodei, R. Anubhai, E. Battenberg, C. Case, J. Casper, B. Catanzaro, J. Chen, M. Chrzanowski, A. Coates, G. Diamos, et al. Deep Speech 2: End-to-end speech recognition in English and Mandarin. arXiv preprint arXiv:1512.02595, 2015.   
+P. Ashby. Understanding phonetics. Routledge, 2013.   
+J. S. Chung and A. Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016a.   
+J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016b.   
+J. Chung, C. Gulcehre, K. Cho, and Y. Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.   
+M. Cooke, J. Barker, S. Cunningham, and X. Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421–2424, 2006.   
+A. Cruttenden. Gimson’s pronunciation of English. Routledge, 2014.   
+G. E. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks for largevocabulary speech recognition. IEEE Transactions on Audio, Speech, and Language Processing, 20(1): 30–42, 2012.   
+F. DeLand. The story of lip-reading, its genesis and development. 1931.   
+R. D. Easton and M. Basala. Perceptual dominance during lipreading. Perception & Psychophysics, 32(6): 562–570, 1982.   
+E. Ferragne and F. Pellegrino. Formant frequencies of vowels in 13 accents of the british isles. Journal of the International Phonetic Association, 40(01):1–34, 2010.   
+C. G. Fisher. Confusions among visually perceived consonants. Journal of Speech, Language, and Hearing Research, 11(4):796–804, 1968.   
+Y. Fu, S. Yan, and T. S. Huang. Classification and feature extraction by simplexization. IEEE Transactions on Information Forensics and Security, 3(1):91–100, 2008.   
+A. Garg, J. Noyola, and S. Bagadia. Lip reading using CNN and LSTM. Technical report, Stanford University, CS231n project report, 2016.   
+S. Gergen, S. Zeiler, A. H. Abdelaziz, R. Nickel, and D. Kolossa. Dynamic stream weighting for turbodecoding-based audiovisual ASR. In Interspeech, pp. 2135–2139, 2016.   
+A. J. Goldschen, O. N. Garcia, and E. D. Petajan. Continuous automatic speech recognition by lipreading. In Motion-Based recognition, pp. 321–343. Springer, 1997.   
+A. Graves and N. Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International Conference on Machine Learning, pp. 1764–1772, 2014.   
+A. Graves and J. Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 18(5):602–610, 2005.   
+A. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber. Connectionist temporal classification: labelling ´ unsegmented sequence data with recurrent neural networks. In ICML, pp. 369–376, 2006.   
+M. Gurban and J.-P. Thiran. Information theoretic feature extraction for audio-visual speech recognition. IEEE Transactions on Signal Processing, 57(12):4765–4776, 2009.   
+K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In IEEE International Conference on Computer Vision, pp. 1026–1034, 2015.   
+S. Hilder, R. Harvey, and B.-J. Theobald. Comparison of human and machine-based lip-reading. In AVSP, pp. 86–89, 2009.   
+G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012.   
+S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.   
+D. Hu, X. Li, et al. Temporal multimodal learning in audiovisual speech recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 3574–3582, 2016.   
+S. Ji, W. Xu, M. Yang, and K. Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221–231, 2013.   
+A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp. 1725–1732, 2014.   
+D. E. King. Dlib-ml: A machine learning toolkit. JMLR, 10(Jul):1755–1758, 2009.   
+D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.   
+O. Koller, H. Ney, and R. Bowden. Deep learning of mouth shapes for sign language. In ICCV Workshop on Assistive Computer Vision and Robotics, pp. 85–91, 2015.   
+A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.   
+P. Lucey and S. Sridharan. Patch-based representation of visual speech. In HCSNet workshop on use of vision in human-computer interaction, pp. 79–85, 2006.   
+A. L. Maas, Z. Xie, D. Jurafsky, and A. Y. Ng. Lexicon-free conversational speech recognition with neural networks. In NAACL, 2015.   
+I. Matthews, T. F. Cootes, J. A. Bangham, S. Cox, and R. Harvey. Extraction of visual features for lipreading. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(2):198–213, 2002.   
+H. McGurk and J. MacDonald. Hearing lips and seeing voices. Nature, 264:746–748, 1976.   
+C. Neti, G. Potamianos, J. Luettin, I. Matthews, H. Glotin, D. Vergyri, J. Sison, and A. Mashari. Audio visual speech recognition. Technical report, IDIAP, 2000.   
+J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng. Multimodal deep learning. In International Conference on Machine Learning, pp. 689–696, 2011.   
+H. Ninomiya, N. Kitaoka, S. Tamura, Y. Iribe, and K. Takeda. Integration of deep bottleneck features for audio-visual speech recognition. In International Speech Communication Association, 2015.   
+K. Noda, Y. Yamaguchi, K. Nakadai, H. G. Okuno, and T. Ogata. Lipreading using convolutional neural network. In INTERSPEECH, pp. 1149–1153, 2014.   
+G. Papandreou, A. Katsamanis, V. Pitsikalis, and P. Maragos. Multimodal fusion and learning with uncertain features applied to audiovisual speech recognition. In Workshop on Multimedia Signal Processing, pp. 264– 267, 2007.   
+G. Papandreou, A. Katsamanis, V. Pitsikalis, and P. Maragos. Adaptive multimodal fusion by uncertainty compensation with application to audiovisual speech recognition. IEEE Transactions on Audio, Speech, and Language Processing, 17(3):423–435, 2009.   
+S. Petridis and M. Pantic. Deep complementary bottleneck features for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2304–2308. IEEE, 2016.   
+V. Pitsikalis, A. Katsamanis, G. Papandreou, and P. Maragos. Adaptive multimodal fusion by uncertainty compensation. In Interspeech, 2006.   
+G. Potamianos, C. Neti, G. Gravier, A. Garg, and A. W. Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9):1306–1326, 2003.   
+C. Sagonas, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In IEEE International Conference on Computer Vision Workshops, pp. 397–403, 2013.   
+K. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.   
+J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for simplicity: The all convolutional net. In ICLR Workshop, 2014.   
+C. Sui, M. Bennamoun, and R. Togneri. Listening with your eyes: Towards a practical visual speech recognition system using deep boltzmann machines. In IEEE International Conference on Computer Vision, pp. 154– 162, 2015.   
+Y. Takashima, R. Aihara, T. Takiguchi, Y. Ariki, N. Mitani, K. Omori, and K. Nakazono. Audio-visual speech recognition using bimodal-trained bottleneck features for a person with severe hearing loss. Interspeech, pp. 277–281, 2016.   
+M. Wand, J. Koutnik, and J. Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 6115–6119, 2016.   
+M. F. Woodward and C. G. Barber. Phoneme perception in lipreading. Journal of Speech, Language, and Hearing Research, 3(3):212–222, 1960.   
+M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pp. 818–833, 2014.   
+G. Zhao, M. Barnard, and M. Pietikainen. Lipreading with local spatiotemporal descriptors. IEEE Transactions on Multimedia, 11(7):1254–1265, 2009.   
+Z. Zhou, G. Zhao, X. Hong, and M. Pietikainen. A review of recent advances in visual speech decoding. ¨ Image and Vision Computing, 32(9):590–605, 2014.
+
+# A ARCHITECTURE DETAILS
+
+In this appendix, we provide additional details about the implementation and architecture.
+
+# A.1 IMPLEMENTATION
+
+LipNet is implemented using Torch, the warp-ctc CTC library (Amodei et al., 2015), and StanfordCTC’s decoder implementation. The network parameters were initialised using He initialisation (He et al., 2015), apart from the square GRU matrices that were orthogonally initialised, as described in (Chung et al., 2014). The models were trained with channel-wise dropout (dropout rate $p = 0 . 5$ ) after each pooling layer and mini-batches of size 50. We used the optimiser Adam (Kingma & Ba, 2014) with a learning rate of $1 0 ^ { - 4 }$ , and the default hyperparameters: a first-moment momentum coefficient of 0.9, a second-moment momentum coefficient of 0.999, and the numerical stability parameter  = 10−8.
+
+The CER and WER scores were computed using CTC beam search with the following parameters for Stanford-CTC’s decoder: beam width 200, $\alpha = 1$ , and $\beta = 1 . 5$ . On top of that, we use a character 5-gram binarised language model, as suggested in (Graves & Jaitly, 2014).
+
+# A.2 LIPNET ARCHITECTURE
+
+The videos were processed with DLib face detector (King, 2009) and the iBug face shape predictor with 68 landmarks (Sagonas et al., 2013). The RGB input frames were normalised using the following per-channel means and standard deviations: $[ \mu _ { R } \bar { = } 0 . 7 1 3 6 , \sigma _ { R } = 0 . 1 1 3 8 , \mu _ { G } = \bar { 0 . 4 9 0 6 } , \sigma _ { G } =$ $0 . 1 0 7 8 , \mu _ { B } = 0 . 3 2 8 3 , \sigma _ { B } = 0 . 0 9 1 7 ]$ ].
+
+Table 3 summarises the LipNet architecture hyperparameters, where $T$ denotes time, $C$ denotes channels, $F$ denotes feature dimension, $H$ and $W$ denote height and width and $V$ denotes the number of words in the vocabulary including the CTC blank symbol.
+
+Table 3: LipNet architecture hyperparameters.   
+
+<table><tr><td>Layer</td><td>Size/Stride/Pad</td><td>Input size</td><td>Dimension order</td></tr><tr><td>STCNN</td><td>3 × 5×5/1,2,2/1,2,2</td><td>75 ×3× 50×100</td><td>T xC×H×W</td></tr><tr><td>Pool STCNN Pool</td><td>1×2× 2/1,2,2 3 × 5× 5/1,2,2/1,2,2 1×2×2/1,2,2</td><td>75 ×32×25× 50 75 × 32 ×12 × 25 × 64×12× 25</td><td>T × C ×H×W T × C×H×1 W</td></tr><tr><td>STCNN</td><td>3 × 3× 3/1,2,2/1,1,1</td><td>75 75 ×64×6 ×12</td><td>T × C×H×W T × C×H×W</td></tr><tr><td>Pool</td><td>1×2× 2/1,2,2</td><td>75 ×96×6×12</td><td>T C×H×W</td></tr><tr><td>Bi-GRU</td><td>256</td><td>75</td><td>×</td></tr><tr><td>Bi-GRU</td><td></td><td>× (96×3×6)</td><td>T × (C ×H×W)</td></tr><tr><td>Linear</td><td>256</td><td>75 × 512</td><td>T×F</td></tr><tr><td>Softmax</td><td>27 +blank</td><td>75 × 512 75×28</td><td>T×F</td></tr></table>
+
+Note that spatiotemporal convolution sizes depend on the number of channels, and the kernel’s three dimensions. Spatiotemporal kernel sizes are specified in the same order as the input size dimensions. The input dimension orderings are given in parentheses in the input size column.
+
+Layers after the Bi-GRU are applied per-timestep.
+
+# A.3 BASELINE-LSTM ARCHITECTURE
+
+Baseline-LSTM replicates the setup of Wand et al. (2016), and is trained the same way as LipNet. The model uses two LSTM layers with 128 neurons. The input frames were converted to grayscale and were down-sampled to $5 0 \times 2 5 \mathrm { p x }$ , dropout $p = 0$ , and the parameters were initialised uniformly with values between $[ - 0 . 0 5 , 0 . 0 5 ]$ .
+
+# B PHONEMES AND VISEMES
+
+Table 4 shows the phoneme to viseme clustering of Neti et al. (2000) and Figure 4 shows LipNet’s full phoneme confusion matrix.
+
+Table 4: Phoneme to viseme clustering of Neti et al. (2000).   
+
+<table><tr><td>Code</td><td>Viseme Class</td><td>Phonemes in Cluster</td></tr><tr><td>V1 V2 V3</td><td>Lip-rounding based vowels</td><td>/ao//ah/ /aa/ ler/ /oyl /aw/ /hh/ /uw/ /uh/ /owl /ael leh/ leyl /ayl</td></tr><tr><td>V4 A</td><td>Alveolar-semivowels</td><td>/ih/ /iy/ /ax/ /V lel/ /r/ lyl</td></tr><tr><td>B</td><td>Alveolar-fricatives</td><td>/s/ /zl</td></tr><tr><td>C</td><td>Alveolar</td><td>/t//d/ /n/len/</td></tr><tr><td>D</td><td>Palato-alveolar</td><td>/sh/ /zh/ /ch/ /jh/</td></tr><tr><td>E</td><td>Bilabial</td><td>/p/ /b/ /m/</td></tr><tr><td>F</td><td>Dental</td><td>/th//dh/</td></tr><tr><td>G</td><td>Labio-dental</td><td>/f/Ivl</td></tr><tr><td>H</td><td>Velar</td><td>/ng/ /k/ Ig/ /wl</td></tr><tr><td>S</td><td>Silence</td><td>/sil/ /sp/</td></tr></table>
+
+![](images/4fcca93a2c8e46800fb40e147a97a1af2435cf77ac1d2220fe4e2d17d3e81497.jpg)  
+Figure 4: LipNet’s full phoneme confusion matrix.

parse/train/BkjLkSqxg/BkjLkSqxg_content_list.json

@@ -0,0 +1,1267 @@
+[
+    {
+        "type": "text",
+        "text": "LIPNET: END-TO-END SENTENCE-LEVEL LIPREADING ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            98,
+            821,
+            121
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yannis M. Assael1,†, Brendan Shillingford1,†, Shimon Whiteson1 & Nando de Freitas1,2,3 ",
+        "bbox": [
+            186,
+            143,
+            799,
+            160
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science, University of Oxford, Oxford, UK   \nGoogle DeepMind, London, UK 2   \nCIFAR, Canada 3   \n{yannis.assael,brendan.shillingford, shimon.whiteson,nando.de.freitas}@cs.ox.ac.uk ",
+        "bbox": [
+            184,
+            160,
+            638,
+            229
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            267,
+            544,
+            282
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lipreading is the task of decoding text from the movement of a speaker’s mouth. Traditional approaches separated the problem into two stages: designing or learning visual features, and prediction. More recent deep lipreading approaches are end-to-end trainable (Wand et al., 2016; Chung & Zisserman, 2016a). However, existing work on models trained end-to-end perform only word classification, rather than sentence-level sequence prediction. Studies have shown that human lipreading performance increases for longer words (Easton & Basala, 1982), indicating the importance of features capturing temporal context in an ambiguous communication channel. Motivated by this observation, we present LipNet, a model that maps a variable-length sequence of video frames to text, making use of spatiotemporal convolutions, a recurrent network, and the connectionist temporal classification loss, trained entirely end-to-end. To the best of our knowledge, LipNet is the first end-to-end sentence-level lipreading model that simultaneously learns spatiotemporal visual features and a sequence model. On the GRID corpus, LipNet achieves $9 5 . 2 \\%$ accuracy in sentence-level, overlapped speaker split task, outperforming experienced human lipreaders and the previous $8 6 . 4 \\%$ word-level state-of-the-art accuracy (Gergen et al., 2016). ",
+        "bbox": [
+            233,
+            300,
+            764,
+            535
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            564,
+            336,
+            580
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lipreading plays a crucial role in human communication and speech understanding, as highlighted by the McGurk effect (McGurk & MacDonald, 1976), where one phoneme’s audio dubbed on top of a video of someone speaking a different phoneme results in a third phoneme being perceived. ",
+        "bbox": [
+            176,
+            597,
+            823,
+            638
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Lipreading is a notoriously difficult task for humans, specially in the absence of context1. Most lipreading actuations, besides the lips and sometimes tongue and teeth, are latent and difficult to disambiguate without context (Fisher, 1968; Woodward & Barber, 1960). For example, Fisher (1968) gives 5 categories of visual phonemes (called visemes), out of a list of 23 initial consonant phonemes, that are commonly confused by people when viewing a speaker’s mouth. Many of these were asymmetrically confused, and observations were similar for final consonant phonemes. ",
+        "bbox": [
+            174,
+            646,
+            825,
+            729
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Consequently, human lipreading performance is poor. Hearing-impaired people achieve an accuracy of only $1 7 \\pm \\mathrm { \\dot { 1 } 2 \\% }$ even for a limited subset of 30 monosyllabic words and $2 1 \\pm 1 1 \\%$ for 30 compound words (Easton & Basala, 1982). An important goal, therefore, is to automate lipreading. Machine lipreaders have enormous practical potential, with applications in improved hearing aids, silent dictation in public spaces, security, speech recognition in noisy environments, biometric identification, and silent-movie processing. ",
+        "bbox": [
+            174,
+            737,
+            825,
+            820
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Machine lipreading is difficult because it requires extracting spatiotemporal features from the video (since both position and motion are important). Recent deep learning approaches attempt to extract those features end-to-end. Most existing work, however, performs only word classification, not sentence-level sequence prediction. ",
+        "bbox": [
+            176,
+            827,
+            823,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we present LipNet, which is to the best of our knowledge, the first end-to-end sentence-level lipreading model. As with modern deep learning based automatic speech recognition (ASR), LipNet is trained end-to-end to make sentence-level predictions. Our model operates at the character-level, using spatiotemporal convolutional neural networks (STCNNs), recurrent neural networks (RNNs), and the connectionist temporal classification loss (CTC) Graves et al. (2006). ",
+        "bbox": [
+            174,
+            103,
+            823,
+            172
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our empirical results on the GRID corpus (Cooke et al., 2006), one of the few public sentence-level datasets, show that LipNet attains a $9 \\bar { 5 } . 2 \\%$ sentence-level word accuracy, in a overlapped speakers split that is popular for benchmarking lipreading methods. The previous best accuracy reported on an aligned word classification version of this task was $8 6 . 4 \\%$ (Gergen et al., 2016). Furthermore, LipNet can generalise across unseen speakers in the GRID corpus with an accuracy of $8 8 . 6 \\%$ . ",
+        "bbox": [
+            174,
+            180,
+            825,
+            251
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We also compare the performance of LipNet with that of hearing-impaired people who can lipread on the GRID corpus task. On average, they achieve an accuracy of ${ \\bar { 5 } } 2 . 3 \\%$ , in contrast to LipNet’s $1 . 6 9 \\times$ higher accuracy in the same sentences. ",
+        "bbox": [
+            174,
+            257,
+            823,
+            299
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Finally, by applying saliency visualisation techniques (Zeiler & Fergus, 2014; Simonyan et al., 2013), we interpret LipNet’s learned behaviour, showing that the model attends to phonologically important regions in the video. Furthermore, by computing intra-viseme and inter-viseme confusion matrices at the phoneme level, we show that almost all of LipNet’s few erroneous predictions occur within visemes, since context is sometimes insufficient for disambiguation. ",
+        "bbox": [
+            174,
+            306,
+            823,
+            376
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 RELATED WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            400,
+            344,
+            416
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we outline various existing approaches to automated lipreading. ",
+        "bbox": [
+            176,
+            435,
+            689,
+            450
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Automated lipreading: Most existing work on lipreading does not employ deep learning. Such work requires either heavy preprocessing of frames to extract image features, temporal preprocessing of frames to extract video features (e.g., optical flow or movement detection), or other types of handcrafted vision pipelines (Matthews et al., 2002; Zhao et al., 2009; Gurban & Thiran, 2009; Papandreou et al., 2007; 2009; Pitsikalis et al., 2006; Lucey & Sridharan, 2006; Papandreou et al., 2009). The automated lipreading literature is too vast to adequately cover, so we refer the reader to Zhou et al. (2014) for an extensive review. ",
+        "bbox": [
+            174,
+            457,
+            825,
+            555
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Notably, Goldschen et al. (1997) were the first to do visual-only sentence-level lipreading using hidden Markov models (HMMs) in a limited dataset, using hand-segmented phones. Later, Neti et al. (2000) were the first to do sentence-level audiovisual speech recognition using an HMM combined with hand-engineered features, on the IBM ViaVoice (Neti et al., 2000) dataset. The authors improve speech recognition performance in noisy environments by fusing visual features with audio ones. The dataset contains 17111 utterances of 261 speakers for training (about 34.9 hours) and is not publicly available. As stated, their visual-only results cannot be interpreted as visual-only recognition, as they are used as rescoring of the noisy audio-only lattices. Using a similar approach, Potamianos et al. (2003) report speaker independent and speaker adapted $9 1 . 6 \\hat { 2 } \\%$ , $8 2 . 3 1 \\%$ WER in the same dataset respectively, and $3 8 . 5 3 \\%$ , $1 6 . 7 7 \\%$ WER in the connected DIGIT corpus, which contains sentences of digits. ",
+        "bbox": [
+            174,
+            563,
+            825,
+            714
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, Gergen et al. (2016) use speaker-dependent training on an LDA-transformed version of the Discrete Cosine Transforms of the mouth regions in an HMM/GMM system. This work holds the previous state-of-the-art on the GRID corpus with a speaker-dependent accuracy of $8 6 . 4 \\%$ . Generalisation across speakers and extraction of motion features is considered an open problem, as noted in (Zhou et al., 2014). LipNet addresses both of these issues. ",
+        "bbox": [
+            174,
+            722,
+            825,
+            791
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Classification with deep learning: In recent years, there have been several attempts to apply deep learning to lipreading. However, all of these approaches perform only word or phoneme classification, whereas LipNet performs full sentence sequence prediction. Approaches include learning multimodal audio-visual representations (Ngiam et al., 2011; Sui et al., 2015; Ninomiya et al., 2015; Petridis & Pantic, 2016), learning visual features as part of a traditional speech-style processing pipeline (e.g. HMMs, GMM-HMMs, etc.) for classifying words and/or phonemes (Almajai et al., 2016; Takashima et al., 2016; Noda et al., 2014; Koller et al., 2015), or combinations thereof (Takashima et al., 2016). Many of these approaches mirror early progress in applying neural networks for acoustic processing in speech recognition (Hinton et al., 2012). ",
+        "bbox": [
+            174,
+            799,
+            825,
+            924
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Chung & Zisserman (2016a) propose spatial and spatiotemporal convolutional neural networks, based on VGG, for word classification. The architectures are evaluated on a word-level dataset BBC TV (333 and 500 classes), but, as reported, their spatiotemporal models fall short of the spatial architectures by an average of around $1 4 \\%$ . Additionally, their models cannot handle variable sequence lengths and they do not attempt sentence-level sequence prediction. ",
+        "bbox": [
+            174,
+            103,
+            825,
+            174
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Chung & Zisserman (2016b) train an audio-visual max-margin matching model for learning pretrained mouth features, which they use as inputs to an LSTM for 10-phrase classification on the OuluVS2 dataset, as well as a non-lipreading task. ",
+        "bbox": [
+            176,
+            180,
+            820,
+            223
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Wand et al. (2016) introduce LSTM recurrent neural networks for lipreading but address neither sentence-level sequence prediction nor speaker independence. ",
+        "bbox": [
+            173,
+            229,
+            823,
+            257
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Garg et al. (2016) apply a VGG pre-trained on faces to classifying words and phrases from the MIRACL-VC1 dataset, which has only 10 words and 10 phrases. However, their best recurrent model is trained by freezing the VGGNet parameters and then training the RNN, rather than training them jointly. Their best model achieves only $5 6 . 0 \\%$ word classification accuracy, and $4 4 . 5 \\%$ phrase classification accuracy, despite both of these being 10-class classification tasks. ",
+        "bbox": [
+            174,
+            263,
+            825,
+            334
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Sequence prediction in speech recognition: The field of automatic speech recognition (ASR) would not be in the state it is today without modern advances in deep learning, many of which have occurred in the context of ASR (Graves et al., 2006; Dahl et al., 2012; Hinton et al., 2012). The connectionist temporal classification loss (CTC) of Graves et al. (2006) drove the movement from deep learning as a component of ASR, to deep ASR systems trained end-to-end (Graves & Jaitly, 2014; Maas et al., 2015; Amodei et al., 2015). As mentioned earlier, much recent lipreading progress has mirrored early progress in ASR, but stopping short of sequence prediction. ",
+        "bbox": [
+            173,
+            342,
+            825,
+            439
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "LipNet is the first end-to-end model that performs sentence-level sequence prediction for visual speech recogntion. That is, we demonstrate the first work that takes as input as sequence of images and outputs a distribution over sequences of tokens; it is trained end-to-end using CTC and thus also does not require alignments. ",
+        "bbox": [
+            174,
+            445,
+            825,
+            502
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Lipreading Datasets: Lipreading datasets (AVICar, AVLetters, AVLetters2, BBC TV, CUAVE, OuluVS1, OuluVS2) are plentiful (Zhou et al., 2014; Chung & Zisserman, 2016a), but most only contain single words or are too small. One exception is the GRID corpus (Cooke et al., 2006), which has audio and video recordings of 34 speakers who produced 1000 sentences each, for a total of 28 hours across 34000 sentences. Table 1 summarises state-of-the-art performance in each of the main lipreading datasets. ",
+        "bbox": [
+            174,
+            508,
+            825,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/f9ed8af4a0acc3a49ffaad3a289b2212ae22a4ac7f09c1f3957847e54ca5f1d1.jpg",
+        "table_caption": [
+            "Table 1: Existing lipreading datasets and the state-of-the-art accuracy reported on these. The size column represents the number of utterances used by the authors for training. Although the GRID corpus contains entire sentences, Gergen et al. (2016) consider only the simpler case of predicting isolated words. LipNet predicts sequences and hence can exploit temporal context to attain much higher accuracy. Phrase-level approaches were treated as plain classification. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Dataset</td><td>Size</td><td>Output</td><td>Accuracy</td></tr><tr><td colspan=\"5\"></td></tr><tr><td>Fu et al. (2008)</td><td>AVICAR</td><td>851</td><td>Digits</td><td>37.9%</td></tr><tr><td>Hu et al. (2016)</td><td>AVLetter</td><td>78</td><td>Alphabet</td><td>64.6%</td></tr><tr><td>Papandreou et al. (2009)</td><td>CUAVE</td><td>1800</td><td>Digits</td><td>83.0%</td></tr><tr><td>Chung &amp; Zisserman (2016a)</td><td>OuluVS1</td><td>200</td><td>Phrases</td><td>91.4%</td></tr><tr><td>Chung &amp; Zisserman (2016b)</td><td>OuluVS2</td><td>520</td><td>Phrases</td><td>94.1%</td></tr><tr><td>Chung &amp; Zisserman ( 1 (2016a)</td><td>BBC TV</td><td>&gt; 400000</td><td>Words</td><td>65.4%</td></tr><tr><td>Gergen et al. (2016)</td><td>GRID</td><td>29700</td><td>Words*</td><td>86.4%</td></tr><tr><td>LipNet</td><td>GRID</td><td>28775</td><td>Sentences</td><td>95.2%</td></tr></table>",
+        "bbox": [
+            225,
+            689,
+            776,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We use the GRID corpus to evaluate LipNet because it is sentence-level and has the most data. The sentences are drawn from the following simple grammar: command $^ { ( 4 ) } + c o l o r ^ { ( 4 ) } +$ prepositio $\\ l _ { \\ l } ^ { ( 4 ) } + l e t t e r ^ { ( 2 5 ) } + d i g i t ^ { ( 1 0 ) } + a d v e r ^ { \\ l } \\ l _ { b } ^ { \\ l ^ { ( 4 ) } }$ , where the number denotes how many word choices there are for each of the 6 word categories. The categories consist of, respectively, {bin, lay, place, set}, {blue, green, red, white}, {at, by, in, with}, $\\{ \\bar { A } , \\ldots , Z \\} \\backslash \\{ W \\}$ , $\\{ \\mathrm { z e r o } , \\ldots , \\mathrm { n i n e } \\}$ , and {again, now, please, $\\operatorname { s o o n } \\}$ , yielding 64000 possible sentences. For example, two sentences in the data are “set blue by A four please” and “place red at C zero again”. ",
+        "bbox": [
+            173,
+            851,
+            825,
+            924
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            103,
+            823,
+            132
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 LIPNET ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            151,
+            271,
+            169
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "LipNet is a neural network architecture for lipreading that maps variable-length sequences of video frames to text sequences, and is trained end-to-end. In this section, we describe LipNet’s building blocks and architecture. ",
+        "bbox": [
+            174,
+            183,
+            825,
+            226
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 SPATIOTEMPORAL CONVOLUTIONS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            242,
+            459,
+            257
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Convolutional neural networks (CNNs), containing stacked convolutions operating spatially over an image, have been instrumental in advancing performance in computer visions tasks such as object recognition that receive an image as input (Krizhevsky et al., 2012). A basic 2D convolution layer from $C$ channels to $C ^ { \\prime }$ channels (without a bias and with unit stride) computes ",
+        "bbox": [
+            173,
+            267,
+            825,
+            325
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/a59b5cbac5113fc2420a6c16897269ce840a7ae8022667cb4ca930c173d234e0.jpg",
+        "text": "$$\n[ \\mathrm { c o n v } ( { \\bf x } , { \\bf w } ) ] _ { c ^ { \\prime } i j } = \\sum _ { c = 1 } ^ { C } \\sum _ { i ^ { \\prime } = 1 } ^ { k _ { w } } \\sum _ { j ^ { \\prime } = 1 } ^ { k _ { h } } w _ { c ^ { \\prime } c i ^ { \\prime } j ^ { \\prime } } x _ { c , i + i ^ { \\prime } , j + j ^ { \\prime } } ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            326,
+            330,
+            669,
+            376
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "for input $\\mathbf { x }$ and weights $\\mathbf { w } \\in \\mathbb { R } ^ { C ^ { \\prime } \\times C \\times k _ { w } \\times k _ { h } }$ where we define $x _ { c i j } = 0$ for $i , j$ out of bounds. Spatiotemporal convolutional neural networks (STCNNs) can process video data by convolving across time, as well as the spatial dimensions (Karpathy et al., 2014; Ji et al., 2013). Hence similarly, ",
+        "bbox": [
+            174,
+            382,
+            825,
+            428
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/4f05ea277e090ca71eea8c536065fda3fffe8a4afea037b608b85bf586cb1f5e.jpg",
+        "text": "$$\n[ \\mathrm { s t c o n v } ( { \\bf x } , { \\bf w } ) ] _ { c ^ { \\prime } t i j } = \\sum _ { c = 1 } ^ { C } \\sum _ { t ^ { \\prime } = 1 } ^ { k _ { t } } \\sum _ { i ^ { \\prime } = 1 } ^ { k _ { w } } \\sum _ { j ^ { \\prime } = 1 } ^ { k _ { h } } w _ { c ^ { \\prime } c t ^ { \\prime } i ^ { \\prime } j ^ { \\prime } } x _ { c , t + t ^ { \\prime } , i + i ^ { \\prime } , j + j ^ { \\prime } } .\n$$",
+        "text_format": "latex",
+        "bbox": [
+            284,
+            433,
+            712,
+            479
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 GATED RECURRENT UNIT",
+        "text_level": 1,
+        "bbox": [
+            174,
+            493,
+            397,
+            508
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Gated Recurrent Unit (GRU) (Chung et al., 2014) is a type of recurrent neural network (RNN) that improves upon earlier RNNs by adding cells and gates for propagating information over more timesteps and learning to control this information flow. It is similar to the Long Short-Term Memory (LSTM) RNN (Hochreiter & Schmidhuber, 1997). We use the standard formulation: ",
+        "bbox": [
+            174,
+            518,
+            825,
+            575
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/aa4e0122cd3eadd52d8ede3bdf70d03d20c3aa91bb24c7867b20f51964fe8a03.jpg",
+        "text": "$$\n\\begin{array} { r l } & { [ \\mathbf { u } _ { t } , \\mathbf { r } _ { t } ] ^ { T } = \\mathrm { s i g m } ( \\mathbf { W } _ { z } \\mathbf { z } _ { t } + \\mathbf { W } _ { h } \\mathbf { h } _ { t - 1 } + \\mathbf { b } _ { g } ) } \\\\ & { \\qquad \\tilde { \\mathbf { h } } _ { t } = \\mathrm { t a n h } ( \\mathbf { U } _ { z } \\mathbf { z } _ { t } + \\mathbf { U } _ { h } ( \\mathbf { r } _ { t } \\odot \\mathbf { h } _ { t - 1 } ) + \\mathbf { b } _ { h } ) } \\\\ & { \\qquad \\mathbf { h } _ { t } = ( \\mathbf { 1 } - \\mathbf { u } _ { t } ) \\odot \\mathbf { h } _ { t - 1 } + \\mathbf { u } _ { t } \\odot \\tilde { \\mathbf { h } } _ { t } } \\end{array}\n$$",
+        "text_format": "latex",
+        "bbox": [
+            336,
+            580,
+            663,
+            643
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $\\mathbf { z } : = \\{ \\mathbf { z } _ { 1 } , \\dots , \\mathbf { z } _ { T } \\}$ is the input sequence to the RNN, $\\odot$ denotes element-wise multiplication, and $\\mathrm { s i g m } ( r ) = 1 / ( 1 + \\exp ( - r ) )$ . We use a bidirectional GRU (Bi-GRU) as introduced by Graves & Schmidhuber (2005) in the context of LSTMs: one RNN maps $\\{ \\mathbf { z } _ { 1 } , \\dotsc , \\mathbf { z } _ { T } \\} \\mapsto \\{ \\overrightarrow { \\mathbf { h } _ { 1 } } , \\dotsc , \\overrightarrow { \\mathbf { h } _ { T } } \\}$ , and another $\\{ \\mathbf { z } _ { T } , \\dotsc , \\mathbf { z } _ { 1 } \\} \\mapsto \\{ { \\overleftarrow { \\mathbf { h } _ { 1 } } } , \\dotsc , { \\overleftarrow { \\mathbf { h } _ { T } } } \\}$ , then $\\mathbf { h } _ { t } : = [ \\overrightarrow { \\mathbf { h } _ { t } } , \\overleftarrow { \\mathbf { h } _ { t } } ]$ . The Bi-GRU ensures that $\\mathbf { h } _ { t }$ depends on $\\mathbf { z } _ { t ^ { \\prime } }$ for all $t ^ { \\prime }$ . To parameterise a distribution over sequences, at time-step $t$ let $p ( u _ { t } | \\mathbf { z } ) =$ softmax $( \\mathrm { m l p } ( \\mathbf { h } _ { t } ; \\mathbf { W } _ { m l p } ) )$ , where mlp is a feed-forward network with weights $\\mathbf { W } _ { m l p }$ . Then we can define the distribution over length- $\\mathcal { T }$ sequences as $\\begin{array} { r } { p ( u _ { 1 } , \\ldots , u _ { T } | \\mathbf { z } ) = \\prod _ { 1 \\leq t \\leq T } p ( u _ { t } | \\mathbf { \\bar { z } } ) } \\end{array}$ , where $T$ is determined by $\\mathbf { z }$ , the input to the GRU. In LipNet, $\\mathbf { z }$ is the output of the STCNN. ",
+        "bbox": [
+            173,
+            646,
+            825,
+            768
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 CONNECTIONIST TEMPORAL CLASSIFICATION ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            784,
+            537,
+            799
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The connectionist temporal classification (CTC) loss (Graves et al., 2006) is widely used in modern speech recognition as it eliminates the need for training data that aligns inputs to target outputs (Amodei et al., 2015; Graves & Jaitly, 2014; Maas et al., 2015). Given a model that outputs a sequence of discrete distributions over the token classes (vocabulary) augmented with a special “blank” token, CTC computes the probability of a sequence by marginalising over all sequences that are defined as equivalent to this sequence. This simultaneously removes the need for alignments and addresses variable-length sequences. Let $V$ denote the set of tokens that the model classifies at a single time-step of its output (vocabulary), and the blank-augmented vocabulary ${ \\tilde { V } } = V \\cup \\{ \\ldots \\}$ where denotes the CTC blank symbol. Define the function $B : \\tilde { V } ^ { * } \\to V ^ { * }$ that, given a string over $\\tilde { V }$ , deletes adjacent duplicate characters and removes blank tokens. For a label sequence $y \\in V ^ { * }$ , CTC defines $\\begin{array} { r } { \\dot { p } ( y | \\mathbf { x } ) = \\sum _ { u \\in \\mathcal { B } ^ { - 1 } ( y ) } } \\end{array}$ s.t. $_ { | u | = T } p \\big ( u _ { 1 } , \\dots , u _ { T } | \\mathbf x \\big )$ , where $T$ is the number of time-steps in the sequence model. For example, if $T = 3$ , CTC defines the probability of a string “am” as $p ( a a m ) \\bar { + } p ( a m m ) + p ( . a m ) + \\bar { p } ( a . m ) + p ( a m . )$ . This sum is computed efficiently by dynamic programming, allowing us to perform maximum likelihood. ",
+        "bbox": [
+            173,
+            809,
+            825,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/1872721e1db4aa3d870c4e29c3ac890f371742e27758f736f3657583d4e43b7e.jpg",
+        "image_caption": [
+            "Figure 1: LipNet architecture. A sequence of $T$ frames is used as input, and is processed by 3 layers of STCNN, each followed by a spatial max-pooling layer. The features extracted are processed by 2 Bi-GRUs; each time-step of the GRU output is processed by a linear layer and a softmax. This end-to-end model is trained with CTC. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            97,
+            826,
+            265
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            173,
+            342,
+            825,
+            431
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.4 LIPNET ARCHITECTURE ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            448,
+            383,
+            462
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 illustrates the LipNet architecture, which starts with $3 \\times$ (spatiotemporal convolutions, channel-wise dropout, spatial max-pooling). Subsequently, the features extracted are followed by two Bi-GRUs. The Bi-GRUs are crucial for efficient further aggregation of the STCNN output. Finally, a linear transformation is applied at each time-step, followed by a softmax over the vocabulary augmented with the CTC blank, and then the CTC loss. All layers use rectified linear unit (ReLU) activation functions. More details including hyperparameters can be found in Table 3 of Appendix A. ",
+        "bbox": [
+            173,
+            473,
+            825,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 LIPREADING EVALUATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            592,
+            424,
+            607
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we evaluate LipNet on the GRID corpus. The augmentation methods employed don’t make use of external data and rely purely on the GRID corpus. ",
+        "bbox": [
+            174,
+            623,
+            823,
+            651
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 DATA AUGMENTATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            667,
+            372,
+            683
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Preprocessing: The GRID corpus consists of 34 subjects, each narrating 1000 sentences. The videos for speaker 21 are missing, and a few others are empty or corrupt, leaving 32746 usable videos. We employ a split (unseen speakers; not previously used in the literature) holding out the data of two male speakers (1 and 2) and two female speakers (20 and 22) for evaluation (3971 videos). The remainder is used for training (28775 videos). We also use a sentence-level variant of the split (overlapped speakers) similar to Wand et al. (2016), where 255 random sentences from each speaker are used for evaluation. All remaining data from all speakers is pooled together for training. All videos are 3 seconds long with a frame rate of 25fps. The videos were processed with the DLib face detector, and the iBug face landmark predictor (Sagonas et al., 2013) with 68 landmarks coupled with an online Kalman Filter. Using these landmarks, we apply an affine transformation to extract a mouth-centred crop of size $1 0 0 \\times 5 0$ pixels per frame. We standardise the RGB channels over the whole training set to have zero mean and unit variance. ",
+        "bbox": [
+            173,
+            694,
+            825,
+            861
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Augmentation: We augment the dataset with simple transformations to reduce overfitting. First, we train on both the regular and the horizontally mirrored image sequence. Second, since the dataset provides word start and end timings for each sentence video, we augment the sentence-level training data with video clips of individual words as additional training instances. These instances have a decay rate of 0.925. Third, to encourage resilience to varying motion speeds by deletion and duplication of frames, this is performed with a per-frame probability of 0.05. The same augmentation methods were followed in all proposed baselines and models. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            103,
+            823,
+            146
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 BASELINES",
+        "text_level": 1,
+        "bbox": [
+            174,
+            161,
+            294,
+            176
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To evaluate LipNet, we compare its performance to that of three hearing-impaired people who can lipread, as well as three ablation models inspired by recent state-of-the-art work (Chung & Zisserman, 2016a; Wand et al., 2016). ",
+        "bbox": [
+            174,
+            188,
+            823,
+            229
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Hearing-Impaired People: This baseline was performed by three members of the Oxford Students’ Disability Community. After being introduced to the grammar of the GRID corpus, they observed 10 minutes of annotated videos from the training dataset, then annotated 300 random videos from the evaluation dataset. When uncertain, they were asked to pick the most probable answer. ",
+        "bbox": [
+            176,
+            237,
+            823,
+            292
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baseline-LSTM: Using the sentence-level training setup of LipNet, we replicate the model architecture of the previous deep learning GRID corpus state-of-the-art (Wand et al., 2016). See Appendix A for more implementation details. ",
+        "bbox": [
+            174,
+            300,
+            823,
+            342
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baseline-2D: Based on the LipNet architecture, we replace the STCNN with spatial-only convolutions similar to those of Chung & Zisserman (2016a). Notably, contrary to the results we observe with LipNet, Chung & Zisserman (2016a) report $1 4 \\%$ and $3 1 \\%$ poorer performance of their STCNNs compared to the 2D architectures in their two datasets. ",
+        "bbox": [
+            174,
+            349,
+            825,
+            405
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baseline-NoLM: Identical to LipNet, but with the language model used in beam search disabled. ",
+        "bbox": [
+            171,
+            412,
+            808,
+            426
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 PERFORMANCE EVALUATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            443,
+            416,
+            457
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To measure the performance of LipNet and the baselines, we compute the word error rate (WER) and the character error rate (CER), standard metrics for the performance of ASR models. We produce approximate maximum-probability predictions from LipNet by performing CTC beam search. WER (or CER) is defined as the minimum number of word (or character) insertions, substitutions, and deletions required to transform the prediction into the ground truth, divided by the number of words (or characters) in the ground truth. Note that WER is usually equal to classification error when the predicted sentence has the same number of words as the ground truth, particularly in our case since almost all errors are substitution errors. ",
+        "bbox": [
+            174,
+            468,
+            825,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2 summarises the performance of LipNet compared to the baselines. According to the literature, the accuracy of human lipreaders is around $2 0 \\%$ (Easton & Basala, 1982; Hilder et al., 2009). As expected, the fixed sentence structure and the limited subset of words for each position in the GRID corpus facilitate the use of context, increasing performance. On the unseen speakers split, the three hearing-impaired people achieve $5 7 . 3 \\%$ , $5 0 . { \\overset { - } { 4 } } { \\overset { - } { \\% } }$ , and $3 5 . 5 \\%$ WER respectively, yielding an average of $4 7 . 7 \\%$ WER. ",
+        "bbox": [
+            174,
+            588,
+            825,
+            670
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/5b324a295a327b8214bbb4da1d271e260d19f3def508ad5771effabb97242ea3.jpg",
+        "table_caption": [
+            "Table 2: Performance of LipNet on the GRID dataset compared to the baselines, measured on two splits: (a) evaluating on only unseen speakers, and (b) evaluating on a 255 video subset of each speakers’ sentences. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Unseen Speakers</td><td colspan=\"2\">Overlapped Speakers</td></tr><tr><td>CER</td><td>WER</td><td>CER</td><td>WER</td></tr><tr><td>Hearing-Impaired Person (avg)</td><td></td><td>47.7%</td><td></td><td></td></tr><tr><td>Baseline-LSTM</td><td>38.4%</td><td>52.8%</td><td>15.2%</td><td>26.3%</td></tr><tr><td>Baseline-2D</td><td>16.2%</td><td>26.7%</td><td>4.3%</td><td>11.6%</td></tr><tr><td>Baseline-NoLM</td><td>6.7%</td><td>13.6%</td><td>2.0%</td><td>5.6%</td></tr><tr><td>LipNet</td><td>6.4%</td><td>11.4%</td><td>1.9%</td><td>4.8%</td></tr></table>",
+        "bbox": [
+            233,
+            742,
+            763,
+            851
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For both unseen and overlapped speakers evaluation, the highest performance is achieved by the architectures enhanced with convolutional stacks. LipNet exhibits a $2 . 3 \\times$ higher performance in the overlapped compared to the unseen speakers split. For unseen speakers, Baseline-2D and LipNet achieve $1 . 8 \\times$ and $4 . 2 \\times$ lower WER, respectively, than hearing-impaired people. ",
+        "bbox": [
+            174,
+            867,
+            825,
+            924
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The WER for unseen speakers Baseline-2D is $2 6 . 7 \\%$ , whereas for LipNet it is $2 . 3 \\times$ lower, at $1 1 . 4 \\%$ . Similarly, the error rate for overlapped speakers was $2 . 4 \\times$ lower for LipNet compared to Baseline2D. Both results demonstrate the importance of combining STCNNs with RNNs. This performance difference confirms the intuition that extracting spatiotemporal features using a STCNN is better than aggregating spatial-only features. This observation contrasts with the empirical observations of Chung & Zisserman (2016a). Furthermore, LipNet’s use of STCNN, RNNs, and CTC cleanly allow processing both variable-length input and variable-length output sequences, whereas the architectures of Chung & Zisserman (2016a) and Chung & Zisserman (2016b) only handle the former. ",
+        "bbox": [
+            173,
+            103,
+            825,
+            215
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Baseline-LSTM exhibits the lowest performance, in both unseen and overlapped speakers, with $5 2 . 8 \\%$ and $2 6 . 3 \\%$ WER, respectively. Interestingly, although Baseline-LSTM replicates the architecture of Wand et al. (2016), and despite the numerous data augmentation methods, the model performs $1 . 3 \\times$ lower than the reported $7 9 . 6 \\%$ word-level accuracy illustrating the difficulty of a sentence-level task even in a restricted grammar. ",
+        "bbox": [
+            174,
+            222,
+            825,
+            291
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Finally, by disabling the language model, the Baseline-NoLM exhibits approximately $1 . 2 \\times$ higher WER than our proposed model. ",
+        "bbox": [
+            174,
+            299,
+            821,
+            327
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 LEARNED REPRESENTATIONS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            348,
+            416,
+            361
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we analyse the learned representations of LipNet from a phonological perspective. First, we create saliency visualisations (Simonyan et al., 2013; Zeiler & Fergus, 2014) to illustrate where LipNet has learned to attend. In particular, we feed an input into the model and greedily decode an output sequence, yielding a CTC alignment $\\hat { u } \\in \\tilde { V } ^ { * }$ (following the notation of Sections 3.2 and 3.3). Then, we compute the gradient of $\\textstyle \\sum _ { t } p ( { \\hat { u } } _ { t } | \\mathbf { x } )$ with respect to the input video frame sequence, but unlike Simonyan et al. (2013), we use guided backpropagation (Springenberg et al., 2014). Second, we train LipNet to predict ARPAbet phonemes, instead of characters, to analyse visual phoneme similarities using intra-viseme and inter-viseme confusion matrices. ",
+        "bbox": [
+            174,
+            375,
+            825,
+            488
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4.1 SALIENCY MAPS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            506,
+            346,
+            521
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We apply saliency visualisation techniques to interpret LipNet’s learned behaviour, showing that the model attends to phonologically important regions in the video. In particular, in Figure 2 we analyse two saliency visualisations for the words please and lay for speaker 25, based on Ashby (2013). ",
+        "bbox": [
+            174,
+            532,
+            825,
+            574
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/d6845168838f936d340381ba209455b7c18df7f8eeafe691226d8b5dedc3de4b.jpg",
+        "image_caption": [
+            "Figure 2: Saliency maps for the words (a) please and (b) lay, produced by backpropagation to the input, showing the places where LipNet has learned to attend. The pictured transcription is given by greedy CTC decoding. CTC blanks are denoted by $\\cdot \\_$ . "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            174,
+            587,
+            825,
+            727
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The production of the word please requires a great deal of articulatory movement at the beginning: the lips are pressed firmly together for the bilabial plosive $/ \\mathrm { p } /$ (frame 1). At the same time, the blade of the tongue comes in contact with the alveolar ridge in anticipation of the following lateral $/ \\mathrm { { l } } / .$ . The lips then part, allowing the compressed air to escape between the lips (frame 2). The jaw and lips then open further, seen in the distance between the midpoints of the upper and lower lips, and the lips spread (increasing the distance between the corners of the mouth), for the close vowel /iy/ (frame 3–4). Since this is a relatively steady-state vowel, lip position remains unchanged for the rest of its duration (frames 4–8), where the attention level drops considerably. The jaw and the lips then close slightly, as the blade of the tongue needs to be brought close to the alveolar ridge, for $/ \\mathbf { Z } /$ (frames 9–10), where attention resumes. ",
+        "bbox": [
+            174,
+            785,
+            825,
+            924
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Lay is interesting since the bulk of frontally visible articulatory movement involves the blade of the tongue coming into contact with the alveolar ridge for $/ \\mathrm { l } / $ (frames 2–6), and then going down for the vowel /ey/ (frames 7–9). That is exactly where most of LipNet’s attention is focused, as there is little change in lip position. ",
+        "bbox": [
+            174,
+            103,
+            825,
+            160
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.4.2 VISEMES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            178,
+            292,
+            193
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "According to DeLand (1931) and Fisher (1968), Alexander Graham Bell first hypothesised that multiple phonemes may be visually identical on a given speaker. This was later verified, giving rise to the concept of a viseme, a visual equivalent of a phoneme (Woodward & Barber, 1960; Fisher, 1968). For our analysis, we use the phoneme-to-viseme mapping of Neti et al. (2000), clustering the phonemes into the following categories: Lip-rounding based vowels (V), Alveolar-semivowels (A), Alveolar-fricatives (B), Alveolar (C), Palato-alveolar (D), Bilabial (E), Dental (F), Labio-dental (G), and Velar (H). The full mapping can be found in Table 4 in Appendix A. The GRID corpus contain 31 out of the 39 phonemes in ARPAbet. We compute confusion matrices between phonemes and then group phonemes into viseme clusters, following Neti et al. (2000). Figure 3 shows the confusion matrices of the 3 most confused viseme categories, as well as the confusions between the viseme categories. The full phoneme confusion matrix is in Figure 4 in Appendix B. ",
+        "bbox": [
+            173,
+            204,
+            825,
+            316
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3feedcc8be2ef8745cfe831ae3909203f4444d251cb9705a44e0d9f421116203.jpg",
+        "image_caption": [
+            "Figure 3: Intra-viseme and inter-viseme confusion matrices, depicting the three categories with the most confusions, as well as the confusions between viseme clusters. Colours are row-normalised to emphasise the errors. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            197,
+            330,
+            799,
+            465
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            523,
+            825,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Given that the speakers are British, the confusion between /aa/ and /ay/ (Figure 3a) is most probably due to the fact that the first element, and the greater part, of the diphthong /ay/ is articulatorily identical with /aa/: an open back unrounded vowel (Ferragne & Pellegrino, 2010). The confusion of /ih/ (a rather close vowel) and /ae/ (a very open vowel) is at first glance surprising, but in fact in the sample /ae/ occurs only in the word at, which is a function word normally pronounced with a reduced, weak vowel /ah/. /ah/ and $/ \\mathrm { i h } I$ are the most frequent unstressed vowels and there is a good deal of variation within and between them, e.g. private and watches (Cruttenden, 2014). ",
+        "bbox": [
+            174,
+            571,
+            825,
+            671
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The confusion within the categories of bilabial stops $/ { \\mathrm { p } } { \\mathrm { ~ b ~ m } } ^ { \\prime }$ and alveolar stops /t d n/ (Figures 3b-c) is unsurprising: complete closure at the same place of articulation makes them look practically identical. The differences of velum action and vocal fold vibration are unobservable from the front. ",
+        "bbox": [
+            176,
+            678,
+            821,
+            719
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Finally, the quality of the viseme categorisation of Neti et al. (2000) is confirmed by the fact that the matrix in Figure 3d is diagonal, with only minor confusion between alveolar (C) and palatoalveolar (D) visemes. Articulatorily, alveolar $/ s \\ z /$ and palato-alveolar $/ \\mathrm { s h \\ z h } /$ fricatives are distinguished by only a small difference in tongue position: against the palate just behind the alveolar ridge, which is not easily observed from the front. The same can be said about dental /th/ and alveolar $/ \\mathrm { t } / $ . ",
+        "bbox": [
+            174,
+            726,
+            823,
+            810
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 CONCLUSIONS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            833,
+            328,
+            849
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We proposed LipNet, the first model to apply deep learning to end-to-end learning of a model that maps sequences of image frames of a speaker’s mouth to entire sentences. The end-to-end model eliminates the need to segment videos into words before predicting a sentence. LipNet requires neither hand-engineered spatiotemporal visual features nor a separately-trained sequence model. ",
+        "bbox": [
+            174,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Our empirical evaluation illustrates the importance of spatiotemporal feature extraction and efficient temporal aggregation, confirming the intuition of Easton & Basala (1982). Furthermore, LipNet greatly outperforms a human lipreading baseline, exhibiting $4 . 1 \\times$ better performance, and $\\bar { 4 } . 8 \\%$ WER which is $2 . 8 \\times$ lower than the word-level state-of-the-art (Gergen et al., 2016) in the GRID corpus. ",
+        "bbox": [
+            174,
+            103,
+            825,
+            172
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While LipNet is already an empirical success, the deep speech recognition literature (Amodei et al., 2015) suggests that performance will only improve with more data. In future work, we hope to demonstrate this by applying LipNet to larger datasets, such as a sentence-level variant of that collected by Chung & Zisserman (2016a). ",
+        "bbox": [
+            174,
+            180,
+            825,
+            236
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Some applications, such as silent dictation, demand the use of video only. However, to extend the range of potential applications of LipNet, we aim to apply this approach to a jointly trained audiovisual speech recognition model, where visual input assists with robustness in noisy environments. ",
+        "bbox": [
+            174,
+            243,
+            825,
+            285
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "ACKNOWLEDGMENTS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            304,
+            326,
+            318
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by an Oxford-Google DeepMind Graduate Scholarship, the EPSRC, and CIFAR. We would also like to thank: NVIDIA for their generous donation of DGX-1 and GTX Titan X GPUs, used in our experiments; Aine Jackson, Brittany Klug and Samantha Pugh for helping ´ us measure the experienced lipreader baseline; Mitko Sabev for his phonetics guidance; Odysseas Votsis for his video production help; and Alex Graves and Oiwi Parker Jones for helpful comments. ",
+        "bbox": [
+            174,
+            328,
+            825,
+            400
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            424,
+            285,
+            439
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "I. Almajai, S. Cox, R. Harvey, and Y. Lan. Improved speaker independent lip reading using speaker adaptive training and deep neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 2722–2726, 2016.   \nD. Amodei, R. Anubhai, E. Battenberg, C. Case, J. Casper, B. Catanzaro, J. Chen, M. Chrzanowski, A. Coates, G. Diamos, et al. Deep Speech 2: End-to-end speech recognition in English and Mandarin. arXiv preprint arXiv:1512.02595, 2015.   \nP. Ashby. Understanding phonetics. Routledge, 2013.   \nJ. S. Chung and A. Zisserman. Lip reading in the wild. In Asian Conference on Computer Vision, 2016a.   \nJ. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Workshop on Multi-view Lip-reading, ACCV, 2016b.   \nJ. Chung, C. Gulcehre, K. Cho, and Y. Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint arXiv:1412.3555, 2014.   \nM. Cooke, J. Barker, S. Cunningham, and X. Shao. An audio-visual corpus for speech perception and automatic speech recognition. The Journal of the Acoustical Society of America, 120(5):2421–2424, 2006.   \nA. Cruttenden. Gimson’s pronunciation of English. Routledge, 2014.   \nG. E. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks for largevocabulary speech recognition. IEEE Transactions on Audio, Speech, and Language Processing, 20(1): 30–42, 2012.   \nF. DeLand. The story of lip-reading, its genesis and development. 1931.   \nR. D. Easton and M. Basala. Perceptual dominance during lipreading. Perception & Psychophysics, 32(6): 562–570, 1982.   \nE. Ferragne and F. Pellegrino. Formant frequencies of vowels in 13 accents of the british isles. Journal of the International Phonetic Association, 40(01):1–34, 2010.   \nC. G. Fisher. Confusions among visually perceived consonants. Journal of Speech, Language, and Hearing Research, 11(4):796–804, 1968.   \nY. Fu, S. Yan, and T. S. Huang. Classification and feature extraction by simplexization. IEEE Transactions on Information Forensics and Security, 3(1):91–100, 2008.   \nA. Garg, J. Noyola, and S. Bagadia. Lip reading using CNN and LSTM. Technical report, Stanford University, CS231n project report, 2016.   \nS. Gergen, S. Zeiler, A. H. Abdelaziz, R. Nickel, and D. Kolossa. Dynamic stream weighting for turbodecoding-based audiovisual ASR. In Interspeech, pp. 2135–2139, 2016.   \nA. J. Goldschen, O. N. Garcia, and E. D. Petajan. Continuous automatic speech recognition by lipreading. In Motion-Based recognition, pp. 321–343. Springer, 1997.   \nA. Graves and N. Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In International Conference on Machine Learning, pp. 1764–1772, 2014.   \nA. Graves and J. Schmidhuber. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 18(5):602–610, 2005.   \nA. Graves, S. Fernandez, F. Gomez, and J. Schmidhuber. Connectionist temporal classification: labelling ´ unsegmented sequence data with recurrent neural networks. In ICML, pp. 369–376, 2006.   \nM. Gurban and J.-P. Thiran. Information theoretic feature extraction for audio-visual speech recognition. IEEE Transactions on Signal Processing, 57(12):4765–4776, 2009.   \nK. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In IEEE International Conference on Computer Vision, pp. 1026–1034, 2015.   \nS. Hilder, R. Harvey, and B.-J. Theobald. Comparison of human and machine-based lip-reading. In AVSP, pp. 86–89, 2009.   \nG. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012.   \nS. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.   \nD. Hu, X. Li, et al. Temporal multimodal learning in audiovisual speech recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pp. 3574–3582, 2016.   \nS. Ji, W. Xu, M. Yang, and K. Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221–231, 2013.   \nA. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pp. 1725–1732, 2014.   \nD. E. King. Dlib-ml: A machine learning toolkit. JMLR, 10(Jul):1755–1758, 2009.   \nD. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.   \nO. Koller, H. Ney, and R. Bowden. Deep learning of mouth shapes for sign language. In ICCV Workshop on Assistive Computer Vision and Robotics, pp. 85–91, 2015.   \nA. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012.   \nP. Lucey and S. Sridharan. Patch-based representation of visual speech. In HCSNet workshop on use of vision in human-computer interaction, pp. 79–85, 2006.   \nA. L. Maas, Z. Xie, D. Jurafsky, and A. Y. Ng. Lexicon-free conversational speech recognition with neural networks. In NAACL, 2015.   \nI. Matthews, T. F. Cootes, J. A. Bangham, S. Cox, and R. Harvey. Extraction of visual features for lipreading. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(2):198–213, 2002.   \nH. McGurk and J. MacDonald. Hearing lips and seeing voices. Nature, 264:746–748, 1976.   \nC. Neti, G. Potamianos, J. Luettin, I. Matthews, H. Glotin, D. Vergyri, J. Sison, and A. Mashari. Audio visual speech recognition. Technical report, IDIAP, 2000.   \nJ. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng. Multimodal deep learning. In International Conference on Machine Learning, pp. 689–696, 2011.   \nH. Ninomiya, N. Kitaoka, S. Tamura, Y. Iribe, and K. Takeda. Integration of deep bottleneck features for audio-visual speech recognition. In International Speech Communication Association, 2015.   \nK. Noda, Y. Yamaguchi, K. Nakadai, H. G. Okuno, and T. Ogata. Lipreading using convolutional neural network. In INTERSPEECH, pp. 1149–1153, 2014.   \nG. Papandreou, A. Katsamanis, V. Pitsikalis, and P. Maragos. Multimodal fusion and learning with uncertain features applied to audiovisual speech recognition. In Workshop on Multimedia Signal Processing, pp. 264– 267, 2007.   \nG. Papandreou, A. Katsamanis, V. Pitsikalis, and P. Maragos. Adaptive multimodal fusion by uncertainty compensation with application to audiovisual speech recognition. IEEE Transactions on Audio, Speech, and Language Processing, 17(3):423–435, 2009.   \nS. Petridis and M. Pantic. Deep complementary bottleneck features for visual speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 2304–2308. IEEE, 2016.   \nV. Pitsikalis, A. Katsamanis, G. Papandreou, and P. Maragos. Adaptive multimodal fusion by uncertainty compensation. In Interspeech, 2006.   \nG. Potamianos, C. Neti, G. Gravier, A. Garg, and A. W. Senior. Recent advances in the automatic recognition of audiovisual speech. Proceedings of the IEEE, 91(9):1306–1326, 2003.   \nC. Sagonas, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic. 300 faces in-the-wild challenge: The first facial landmark localization challenge. In IEEE International Conference on Computer Vision Workshops, pp. 397–403, 2013.   \nK. Simonyan, A. Vedaldi, and A. Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013.   \nJ. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller. Striving for simplicity: The all convolutional net. In ICLR Workshop, 2014.   \nC. Sui, M. Bennamoun, and R. Togneri. Listening with your eyes: Towards a practical visual speech recognition system using deep boltzmann machines. In IEEE International Conference on Computer Vision, pp. 154– 162, 2015.   \nY. Takashima, R. Aihara, T. Takiguchi, Y. Ariki, N. Mitani, K. Omori, and K. Nakazono. Audio-visual speech recognition using bimodal-trained bottleneck features for a person with severe hearing loss. Interspeech, pp. 277–281, 2016.   \nM. Wand, J. Koutnik, and J. Schmidhuber. Lipreading with long short-term memory. In IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 6115–6119, 2016.   \nM. F. Woodward and C. G. Barber. Phoneme perception in lipreading. Journal of Speech, Language, and Hearing Research, 3(3):212–222, 1960.   \nM. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, pp. 818–833, 2014.   \nG. Zhao, M. Barnard, and M. Pietikainen. Lipreading with local spatiotemporal descriptors. IEEE Transactions on Multimedia, 11(7):1254–1265, 2009.   \nZ. Zhou, G. Zhao, X. Hong, and M. Pietikainen. A review of recent advances in visual speech decoding. ¨ Image and Vision Computing, 32(9):590–605, 2014. ",
+        "bbox": [
+            171,
+            445,
+            826,
+            926
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            89,
+            828,
+            904
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            171,
+            99,
+            826,
+            554
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A ARCHITECTURE DETAILS ",
+        "text_level": 1,
+        "bbox": [
+            178,
+            102,
+            421,
+            118
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "In this appendix, we provide additional details about the implementation and architecture. ",
+        "bbox": [
+            171,
+            133,
+            759,
+            148
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 IMPLEMENTATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            167,
+            344,
+            183
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "LipNet is implemented using Torch, the warp-ctc CTC library (Amodei et al., 2015), and StanfordCTC’s decoder implementation. The network parameters were initialised using He initialisation (He et al., 2015), apart from the square GRU matrices that were orthogonally initialised, as described in (Chung et al., 2014). The models were trained with channel-wise dropout (dropout rate $p = 0 . 5$ ) after each pooling layer and mini-batches of size 50. We used the optimiser Adam (Kingma & Ba, 2014) with a learning rate of $1 0 ^ { - 4 }$ , and the default hyperparameters: a first-moment momentum coefficient of 0.9, a second-moment momentum coefficient of 0.999, and the numerical stability parameter \u000f = 10−8. ",
+        "bbox": [
+            173,
+            194,
+            825,
+            306
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The CER and WER scores were computed using CTC beam search with the following parameters for Stanford-CTC’s decoder: beam width 200, $\\alpha = 1$ , and $\\beta = 1 . 5$ . On top of that, we use a character 5-gram binarised language model, as suggested in (Graves & Jaitly, 2014). ",
+        "bbox": [
+            176,
+            313,
+            825,
+            356
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 LIPNET ARCHITECTURE ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            375,
+            388,
+            388
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The videos were processed with DLib face detector (King, 2009) and the iBug face shape predictor with 68 landmarks (Sagonas et al., 2013). The RGB input frames were normalised using the following per-channel means and standard deviations: $[ \\mu _ { R } \\bar { = } 0 . 7 1 3 6 , \\sigma _ { R } = 0 . 1 1 3 8 , \\mu _ { G } = \\bar { 0 . 4 9 0 6 } , \\sigma _ { G } =$ $0 . 1 0 7 8 , \\mu _ { B } = 0 . 3 2 8 3 , \\sigma _ { B } = 0 . 0 9 1 7 ]$ ]. ",
+        "bbox": [
+            174,
+            400,
+            823,
+            458
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 3 summarises the LipNet architecture hyperparameters, where $T$ denotes time, $C$ denotes channels, $F$ denotes feature dimension, $H$ and $W$ denote height and width and $V$ denotes the number of words in the vocabulary including the CTC blank symbol. ",
+        "bbox": [
+            174,
+            464,
+            821,
+            506
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/73f9f065b96cbe38f10f260e5c8ba0ad5df5cfe03b3df0673857386f45498514.jpg",
+        "table_caption": [
+            "Table 3: LipNet architecture hyperparameters. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Layer</td><td>Size/Stride/Pad</td><td>Input size</td><td>Dimension order</td></tr><tr><td>STCNN</td><td>3 × 5×5/1,2,2/1,2,2</td><td>75 ×3× 50×100</td><td>T xC×H×W</td></tr><tr><td>Pool STCNN Pool</td><td>1×2× 2/1,2,2 3 × 5× 5/1,2,2/1,2,2 1×2×2/1,2,2</td><td>75 ×32×25× 50 75 × 32 ×12 × 25 × 64×12× 25</td><td>T × C ×H×W T × C×H×1 W</td></tr><tr><td>STCNN</td><td>3 × 3× 3/1,2,2/1,1,1</td><td>75 75 ×64×6 ×12</td><td>T × C×H×W T × C×H×W</td></tr><tr><td>Pool</td><td>1×2× 2/1,2,2</td><td>75 ×96×6×12</td><td>T C×H×W</td></tr><tr><td>Bi-GRU</td><td>256</td><td>75</td><td>×</td></tr><tr><td>Bi-GRU</td><td></td><td>× (96×3×6)</td><td>T × (C ×H×W)</td></tr><tr><td>Linear</td><td>256</td><td>75 × 512</td><td>T×F</td></tr><tr><td>Softmax</td><td>27 +blank</td><td>75 × 512 75×28</td><td>T×F</td></tr></table>",
+        "bbox": [
+            214,
+            549,
+            782,
+            737
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Note that spatiotemporal convolution sizes depend on the number of channels, and the kernel’s three dimensions. Spatiotemporal kernel sizes are specified in the same order as the input size dimensions. The input dimension orderings are given in parentheses in the input size column. ",
+        "bbox": [
+            176,
+            758,
+            823,
+            801
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Layers after the Bi-GRU are applied per-timestep. ",
+        "bbox": [
+            174,
+            808,
+            501,
+            821
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.3 BASELINE-LSTM ARCHITECTURE ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            840,
+            457,
+            854
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Baseline-LSTM replicates the setup of Wand et al. (2016), and is trained the same way as LipNet. The model uses two LSTM layers with 128 neurons. The input frames were converted to grayscale and were down-sampled to $5 0 \\times 2 5 \\mathrm { p x }$ , dropout $p = 0$ , and the parameters were initialised uniformly with values between $[ - 0 . 0 5 , 0 . 0 5 ]$ . ",
+        "bbox": [
+            176,
+            867,
+            823,
+            924
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B PHONEMES AND VISEMES",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            428,
+            118
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 4 shows the phoneme to viseme clustering of Neti et al. (2000) and Figure 4 shows LipNet’s full phoneme confusion matrix. ",
+        "bbox": [
+            173,
+            133,
+            825,
+            161
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/75c522692c65d2d1ca6adf0be67397fee328b03b9f3bac063360bfa73afe49b0.jpg",
+        "table_caption": [
+            "Table 4: Phoneme to viseme clustering of Neti et al. (2000). "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Code</td><td>Viseme Class</td><td>Phonemes in Cluster</td></tr><tr><td>V1 V2 V3</td><td>Lip-rounding based vowels</td><td>/ao//ah/ /aa/ ler/ /oyl /aw/ /hh/ /uw/ /uh/ /owl /ael leh/ leyl /ayl</td></tr><tr><td>V4 A</td><td>Alveolar-semivowels</td><td>/ih/ /iy/ /ax/ /V lel/ /r/ lyl</td></tr><tr><td>B</td><td>Alveolar-fricatives</td><td>/s/ /zl</td></tr><tr><td>C</td><td>Alveolar</td><td>/t//d/ /n/len/</td></tr><tr><td>D</td><td>Palato-alveolar</td><td>/sh/ /zh/ /ch/ /jh/</td></tr><tr><td>E</td><td>Bilabial</td><td>/p/ /b/ /m/</td></tr><tr><td>F</td><td>Dental</td><td>/th//dh/</td></tr><tr><td>G</td><td>Labio-dental</td><td>/f/Ivl</td></tr><tr><td>H</td><td>Velar</td><td>/ng/ /k/ Ig/ /wl</td></tr><tr><td>S</td><td>Silence</td><td>/sil/ /sp/</td></tr></table>",
+        "bbox": [
+            264,
+            205,
+            735,
+            419
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/4fcca93a2c8e46800fb40e147a97a1af2435cf77ac1d2220fe4e2d17d3e81497.jpg",
+        "image_caption": [
+            "Figure 4: LipNet’s full phoneme confusion matrix. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            245,
+            446,
+            758,
+            842
+        ],
+        "page_idx": 12
+    }
+]

parse/train/S1PWi_lC-/S1PWi_lC-.md

@@ -0,0 +1,132 @@
+# MULTI-TASK LEARNING ON MNIST IMAGE DATASETS
+
+Anonymous authors Paper under double-blind review
+
+# ABSTRACT
+
+We apply multi-task learning to image classification tasks on MNIST-like datasets. MNIST dataset has been referred to as the drosophila of machine learning and has been the testbed of many learning theories. The NotMNIST dataset and the FashionMNIST dataset have been created with the MNIST dataset as reference. In this work, we exploit these MNIST-like datasets for multi-task learning. The datasets are pooled together for learning the parameters of joint classification networks. Then the learned parameters are used as the initial parameters to retrain disjoint classification networks. The baseline recognition model are all-convolution neural networks. Without multi-task learning, the recognition accuracies for MNIST, NotMNIST and FashionMNIST are $9 9 . 5 6 \%$ , $9 7 . 2 2 \%$ and $9 4 . 3 2 \%$ respectively. With multi-task learning to pre-train the networks, the recognition accuracies are respectively $9 9 . 7 0 \%$ , $9 7 . 4 6 \%$ and $9 5 . 2 5 \%$ . The results re-affirm that multi-task learning framework, even with data with different genres, does lead to significant improvement.
+
+# 1 INTRODUCTION
+
+Multi-task learning (Caruana (1998)) enjoys the idea of pooling information that can be learned from data collected for multiple related tasks. Multiple sources of information can stem from multiple datasets, or even a single dataset, for multiple tasks. In this work, we focus on the case of using multiple datasets for multiple tasks. Namely, we use MNIST, FashionMNIST, and NotMNIST image datasets collected for digit recognition, fashion item recognition, and letter recognition, respectively.
+
+Information sharing in multi-task training can be achieved in various formality. For neural-network based deep learning, the sharing can happen at the input layer, the hidden layers, or the output layer. Input-layer multi-tasking combines heterogeneous input data, hidden-layer multi-tasking shares multiple groups of hidden layer units, and output-layer multi-tasking pools multiple output groups of categories. The implementation of a multi-task learning system depends on the data and the tasks at hand.
+
+Multi-task learning has been successfully applied to many applications of machine learning, from natural language processing (Collobert & Weston (2008)) and speech recognition (Deng et al. (2013)) to computer vision (Ren et al. (2015)) and drug discovery (Ramsundar et al. (2015)). A recent review of multi-task learning in deep learning can be found in (Ruder (2017)).
+
+# 2 MNIST DATASETS
+
+# 2.1 MNIST
+
+The MNIST dataset (LeCun et al. (1998)) consists of a training set of 60,000 images, and a test set of 10,000 images. MNIST is often referred to as the drosophila of machine learning, as it is an ideal testbed for new machine learning theories or methods on real-world data. Table 1 lists the state-of-the-art performance on MNIST dataset. A few examples of MNIST dataset are shown in Table 2, alongside with examples of the other MNIST-like datasets.
+
+Table 1: Recognition error rates on MNIST task of state-of-the-art systems   
+Table 2: Image examples of MNIST-like datasets   
+
+<table><tr><td>Method</td><td>Error rate</td></tr><tr><td>Wan et al. (2013)</td><td>0.21%</td></tr><tr><td>Ciregan et al. (2012)</td><td>0.23%</td></tr><tr><td>Sato et al. (2015)</td><td>0.23%</td></tr><tr><td>Chang &amp; Chen (2015)</td><td>0.24%</td></tr></table>
+
+# 2.2 FASHIONMNIST
+
+Xiao et al. (2017) presents the FashionMNIST dataset. It consists of images from the assortment on Zalandos website. As given out by name, the configuration of the FashionMNIST dataset completely parallels the configuration of the MNIST dataset. FashionMNIST consists of a training set of 60,000 images and a test set of 10,000 images. Each image is a $2 8 \times 2 8$ grayscale image associated with a label from 10 classes. FashionMNIST poses a more challenging classification task than the MNIST digits data. A leaderboard for FashionMNIST has been created and maintained at https://github.com/zalandoresearch/fashion-mnist.
+
+![](images/5c5d0bb78bdb2914e2c8106fecbe3c63d4f1cf2198f2c63c498993d15754d522.jpg)
+
+# 2.3 NOTMNIST
+
+Bulatov (2011) presents the NotMIST dataset. NotMNIST dataset consists of more than 500k $2 8 \times 2 8$ greyscale images of English letters from $A$ to $J$ , including a small hand-cleaned subset and a large uncleaned subset. From the uncleaned subset, we randomly select 60,000 examples as train set and 10,000 examples as test set.
+
+# 3 METHOD
+
+In short, we apply multi-task learning to learn from the MNIST-like datasets to pre-train the parameters of the all-convolution neural networks for individual image recognition tasks. The overall framework is depicted in Figure 1.
+
+![](images/dd2a5e432453b5c842b9c8cf4f67007f96f30244bcf996cc624c9f58cdb9c61b.jpg)  
+Figure 1: The proposed multi-task learning framework
+
+# 3.1 ALL-CONVOLUTION NEURAL NETWORKS
+
+A convolutional neural networks (CNN) (LeCun et al. (1998)) usually consists of alternating convolution layers and pooling layers. In contrast, an all-convolution neural network (Springenberg et al. (2014)) replace pooling layers by convolution layers. A feature map f created at a convolution layer can be represented by a three-dimensional tensor of order $W \times H \times N$ , where $W$ and $H$ are the width and height, $N$ is the number of channels. A pooling operation with a size of $k \times k$ and a stride of $r$ applied to a three-dimensional tensor results in another tensor s with
+
+$$
+s _ { i , j , n } = \left( \sum _ { w = - \lfloor k / 2 \rfloor } ^ { \lfloor k / 2 \rfloor } \sum _ { h = - \lfloor k / 2 \rfloor } ^ { \lfloor k / 2 \rfloor } \left. f _ { r \cdot i + w , r \cdot j + h , n } \right. ^ { p } \right) ^ { 1 / p }
+$$
+
+where the first 2 subscripts index position in the map, and the third subscript indexed the channel. If a convolution operation with the same stride were applied to $\mathbf { f }$ , we would have a tensor $\mathbf { c }$ with
+
+$$
+c _ { i , j , o } = \sigma \left( \sum _ { w = - \lfloor k / 2 \rfloor } ^ { \lfloor k / 2 \rfloor } \sum _ { h = - \lfloor k / 2 \rfloor } ^ { \lfloor k / 2 \rfloor } \sum _ { n = 1 } ^ { N } t _ { w , h , n , o } f _ { r \cdot i + w , r \cdot j + h , n } \right)
+$$
+
+where $\mathbf { t }$ is the convolution kernel tensor, and $\sigma ( \cdot )$ is an activation function. Thus, a pooling operation can be seen as a convolution operation with uniform kernel tensor and with $L ^ { p }$ -norm as the activation function.
+
+# 4 EXPERIMENT
+
+# 4.1 CLASSIFIER ARCHITECTURE AND MULTI-TASK LEARNING
+
+The architecture of an all-convolution neural network for a single task is shown in Figure 2. The multi-task learning classifier has the same architecture as a single-task classifier except that the width of the output layer is proportional to the number of tasks. The target label is enhanced accordingly by zero-padding. As mentioned before, we combine the training data together, so that the training label dimension is increased from $[ 1 \times 1 0 ]$ to $[ 1 \times 2 0 ]$ by zero-padding. This is shown in Figure 3.
+
+![](images/ac23d7dabb36a6b95a776544ed2d7e602eeee03c2fb47483ca18f15b05732f49.jpg)  
+Figure 2: The architecture of an all-convolution neural network for a single task
+
+![](images/5f04f6f503b7669c78d2b79422ebfa31a22039d12bda9d76aa4b1e35e970d5b8.jpg)  
+Figure 3: The architecture of all-convolution neural network for multi-task learning
+
+Multi-task learning with different combination of datasets are evaluated, namely 3 bi-task learning and 1 tri-task learning. For example, we combine the MNIST train dataset and FashionMNIST train dataset together. Thus, there are 120,000 examples to the bi-task learning network for MNIST and FashionMNIST. Each training example has an input image size of $2 8 \times 2 8$ and an output 1-hot vector size of 20.
+
+Each network is trained with 50 epochs. A two-stage learning rate decay scheme is implemented. The initial learning rate is $1 0 ^ { - 3 }$ for the first stage of 25 epochs, and $1 0 ^ { - 5 }$ for the second stage of 25 epochs. When training a model, it is often recommended to lower the learning rate as the training progresses. So, we let learning rate decay in each epochs. The decay rate is $\frac { \mathbf { \bar { \rho } } _ { 1 } } { 1 + d \times n }$ where $d$ is the learning rate set in the begin divided by 25. $n$ is the number of current epochs. Figure 4 shows the learning rate scheme. The size of a mini-batch is set to 100, and the Adam optimizer is used.
+
+When multi-task learning is complete, the parameters in the network is used to initialize single-task classifiers. Figure 5 illustrates how parameters are handed over. The single-task classifiers are then re-trained to perform their respective classification tasks.
+
+![](images/b29cbf66c659b1a0765c87255110fee722a9145d3f5c6e61aa9e32ecc4b603c9.jpg)  
+Figure 4: The learning rate scheme.
+
+# 4.2 RESULTS
+
+The experimental results are summarized in Table 3. We are happy to see that multi-task learning works, even though the MNIST, NotMNIST, and FashionMNIST datasets are images from totally different classes. The bi-task learning systems are always better than the single-task systems. Furthermore, the tri-task learning systems are the best, except for the MNIST $( 0 . 0 1 \%$ difference). The relative reduction in error rates by tri-task learning are respectively $3 1 . 8 \%$ for MNIST $( 9 9 . 5 6 \%$ to $9 9 . 7 0 \%$ ), $1 6 . 4 \%$ for FashionMNIST $( 9 4 . 3 2 \%$ to $9 5 . 2 5 \%$ ), and $8 . 6 \%$ for NotMNIST $( 9 7 . 2 2 \%$ to $9 7 . 4 6 \%$ ). The results confirm that multi-task learning is able to learn representation which is universal and robust to different tasks.
+
+![](images/d6df547dae357a04f5bf6418d55fbedf27c8ded7027bb51ff7a15333d33358fe.jpg)  
+Figure 5: Passing parameters from multi-task learning network to single-task networks
+
+Table 3: Image recognition accuracy rates for MNIST tasks with multi-task learning   
+
+<table><tr><td>dataset</td><td>single-task</td><td>F+N</td><td>M+F</td><td>M+N</td><td>F+N+M</td></tr><tr><td>MNIST</td><td>99.56%</td><td></td><td>99.71%</td><td>99.70%</td><td>99.70%</td></tr><tr><td>NotMNIST</td><td>97.22%</td><td>97.38%</td><td></td><td>97.40%</td><td>97.46%</td></tr><tr><td>FashionMNIST</td><td>94.32%</td><td>95.20%</td><td>95.18%</td><td></td><td>95.25%</td></tr><tr><td></td><td></td><td></td><td></td><td>1</td><td></td></tr></table>
+
+To better understand the effect of multi-task learning, we plot the distribution of the high dimensional-data of the classes by the $t$ -distributed stochastic neighbor embedding (t-SNE) (van der Maaten & Hinton (2008)). Figure 6 shows the t-SNE of the output values of several hidden layers of the classifiers. The separation between the learned class manifolds appears to increase with the integration of multi-task learning. The representation learned with multi-task learning in the loop looks better indeed.
+
+# 5 CONCLUSION
+
+In this paper, we use multi-task learning in pre-training an all-convolution neural network model. We pass the parameters of trained multi-task models to single-task models. Evaluation on MNISTlike datasets show that using multi-task learning can improve image recognition accuracy. The more data we use, the better results we get. This agrees with statistical learning theory that using more data reduces the generalization gap, thus improving test set performance, even if the data comes from a different domain. The classification tasks of the images of digits, letters, and fashion items share parts of their hierarchical representations. By multi-task learning, it is possible to make such common representation robust to help individual classification tasks.
+
+![](images/3e0390dfaf853045e89c03ded997c69f36c15fe35724b5b063088a7dcb6a5c9f.jpg)  
+Figure 6: Visualization of data manifolds with t-SNE. The left column is the case with multi-task learning, and the right column is the case without multi-task learning.
+
+# REFERENCES
+
+Yaroslav Bulatov. Notmnist dataset. Available at: http://yaroslavvb. blogspot. it/2011/09/notmnistdataset. html, 2011.
+
+Rich Caruana. Multitask learning. In Learning to learn, pp. 95–133. Springer, 1998.
+
+Jia-Ren Chang and Yong-Sheng Chen. Batch-normalized maxout network in network. arXiv preprint arXiv:1511.02583, 2015.
+
+Dan Ciregan, Ueli Meier, and Jurgen Schmidhuber. Multi-column deep neural networks for image ¨ classification. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3642–3649, 2012.
+
+Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th International Conference on Machine Learning, pp. 160–167, 2008.
+
+Li Deng, Geoffrey Hinton, and Brian Kingsbury. New types of deep neural network learning for speech recognition and related applications: An overview. In Proceedings of the IEEE Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8599–8603, 2013.
+
+Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
+
+Bharath Ramsundar, Steven Kearnes, Patrick Riley, Dale Webster, David Konerding, and Vijay Pande. Massively multitask networks for drug discovery. arXiv preprint arXiv:1502.02072, 2015.
+
+Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In Proceedings of the Advances in Neural Information Processing Systems, pp. 91–99, 2015.
+
+Sebastian Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098, 2017.
+
+Ikuro Sato, Hiroki Nishimura, and Kensuke Yokoi. Apac: Augmented pattern classification with neural networks. arXiv preprint arXiv:1505.03229, 2015.
+
+Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014.
+
+L.J.P. van der Maaten and G.E. Hinton. Visualizing high-dimensional data using t-SNE. Machine Learning Research, 9:2579–2605, 2008.
+
+Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. Regularization of neural networks using dropconnect. In Proceedings of the 30th international conference on machine learning (ICML), pp. 1058–1066, 2013.
+
+Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.

parse/train/S1PWi_lC-/S1PWi_lC-_content_list.json

@@ -0,0 +1,728 @@
+[
+    {
+        "type": "text",
+        "text": "MULTI-TASK LEARNING ON MNIST IMAGE DATASETS ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            98,
+            820,
+            121
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Anonymous authors Paper under double-blind review ",
+        "bbox": [
+            183,
+            145,
+            398,
+            172
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            209,
+            544,
+            226
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We apply multi-task learning to image classification tasks on MNIST-like datasets. MNIST dataset has been referred to as the drosophila of machine learning and has been the testbed of many learning theories. The NotMNIST dataset and the FashionMNIST dataset have been created with the MNIST dataset as reference. In this work, we exploit these MNIST-like datasets for multi-task learning. The datasets are pooled together for learning the parameters of joint classification networks. Then the learned parameters are used as the initial parameters to retrain disjoint classification networks. The baseline recognition model are all-convolution neural networks. Without multi-task learning, the recognition accuracies for MNIST, NotMNIST and FashionMNIST are $9 9 . 5 6 \\%$ , $9 7 . 2 2 \\%$ and $9 4 . 3 2 \\%$ respectively. With multi-task learning to pre-train the networks, the recognition accuracies are respectively $9 9 . 7 0 \\%$ , $9 7 . 4 6 \\%$ and $9 5 . 2 5 \\%$ . The results re-affirm that multi-task learning framework, even with data with different genres, does lead to significant improvement. ",
+        "bbox": [
+            233,
+            250,
+            764,
+            444
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            489,
+            336,
+            506
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-task learning (Caruana (1998)) enjoys the idea of pooling information that can be learned from data collected for multiple related tasks. Multiple sources of information can stem from multiple datasets, or even a single dataset, for multiple tasks. In this work, we focus on the case of using multiple datasets for multiple tasks. Namely, we use MNIST, FashionMNIST, and NotMNIST image datasets collected for digit recognition, fashion item recognition, and letter recognition, respectively. ",
+        "bbox": [
+            174,
+            529,
+            823,
+            599
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Information sharing in multi-task training can be achieved in various formality. For neural-network based deep learning, the sharing can happen at the input layer, the hidden layers, or the output layer. Input-layer multi-tasking combines heterogeneous input data, hidden-layer multi-tasking shares multiple groups of hidden layer units, and output-layer multi-tasking pools multiple output groups of categories. The implementation of a multi-task learning system depends on the data and the tasks at hand. ",
+        "bbox": [
+            174,
+            606,
+            823,
+            689
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-task learning has been successfully applied to many applications of machine learning, from natural language processing (Collobert & Weston (2008)) and speech recognition (Deng et al. (2013)) to computer vision (Ren et al. (2015)) and drug discovery (Ramsundar et al. (2015)). A recent review of multi-task learning in deep learning can be found in (Ruder (2017)). ",
+        "bbox": [
+            174,
+            696,
+            825,
+            752
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 MNIST DATASETS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            784,
+            362,
+            801
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2.1 MNIST ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            821,
+            269,
+            837
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The MNIST dataset (LeCun et al. (1998)) consists of a training set of 60,000 images, and a test set of 10,000 images. MNIST is often referred to as the drosophila of machine learning, as it is an ideal testbed for new machine learning theories or methods on real-world data. Table 1 lists the state-of-the-art performance on MNIST dataset. A few examples of MNIST dataset are shown in Table 2, alongside with examples of the other MNIST-like datasets. ",
+        "bbox": [
+            176,
+            853,
+            823,
+            922
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "table",
+        "img_path": "images/e8eeaffd3016938251eb0545e558abe870de0f421566692218898bdc8a977d42.jpg",
+        "table_caption": [
+            "Table 1: Recognition error rates on MNIST task of state-of-the-art systems ",
+            "Table 2: Image examples of MNIST-like datasets "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Error rate</td></tr><tr><td>Wan et al. (2013)</td><td>0.21%</td></tr><tr><td>Ciregan et al. (2012)</td><td>0.23%</td></tr><tr><td>Sato et al. (2015)</td><td>0.23%</td></tr><tr><td>Chang &amp; Chen (2015)</td><td>0.24%</td></tr></table>",
+        "bbox": [
+            367,
+            131,
+            629,
+            218
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 FASHIONMNIST ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            252,
+            333,
+            267
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Xiao et al. (2017) presents the FashionMNIST dataset. It consists of images from the assortment on Zalandos website. As given out by name, the configuration of the FashionMNIST dataset completely parallels the configuration of the MNIST dataset. FashionMNIST consists of a training set of 60,000 images and a test set of 10,000 images. Each image is a $2 8 \\times 2 8$ grayscale image associated with a label from 10 classes. FashionMNIST poses a more challenging classification task than the MNIST digits data. A leaderboard for FashionMNIST has been created and maintained at https://github.com/zalandoresearch/fashion-mnist. ",
+        "bbox": [
+            173,
+            280,
+            825,
+            377
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c5d0bb78bdb2914e2c8106fecbe3c63d4f1cf2198f2c63c498993d15754d522.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            279,
+            422,
+            717,
+            704
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 NOTMNIST ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            739,
+            300,
+            755
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Bulatov (2011) presents the NotMIST dataset. NotMNIST dataset consists of more than 500k $2 8 \\times 2 8$ greyscale images of English letters from $A$ to $J$ , including a small hand-cleaned subset and a large uncleaned subset. From the uncleaned subset, we randomly select 60,000 examples as train set and 10,000 examples as test set. ",
+        "bbox": [
+            174,
+            767,
+            825,
+            824
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 METHOD ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            847,
+            282,
+            863
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In short, we apply multi-task learning to learn from the MNIST-like datasets to pre-train the parameters of the all-convolution neural networks for individual image recognition tasks. The overall framework is depicted in Figure 1. ",
+        "bbox": [
+            176,
+            881,
+            823,
+            924
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd2a5e432453b5c842b9c8cf4f67007f96f30244bcf996cc624c9f58cdb9c61b.jpg",
+        "image_caption": [
+            "Figure 1: The proposed multi-task learning framework "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            192,
+            104,
+            818,
+            426
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 ALL-CONVOLUTION NEURAL NETWORKS ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            486,
+            506,
+            501
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "A convolutional neural networks (CNN) (LeCun et al. (1998)) usually consists of alternating convolution layers and pooling layers. In contrast, an all-convolution neural network (Springenberg et al. (2014)) replace pooling layers by convolution layers. A feature map f created at a convolution layer can be represented by a three-dimensional tensor of order $W \\times H \\times N$ , where $W$ and $H$ are the width and height, $N$ is the number of channels. A pooling operation with a size of $k \\times k$ and a stride of $r$ applied to a three-dimensional tensor results in another tensor s with ",
+        "bbox": [
+            173,
+            512,
+            826,
+            597
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/fdb9bf7614b27f5cc2e538cf67b4be6391455930356ef00fdf9cb1dcf3e88daa.jpg",
+        "text": "$$\ns _ { i , j , n } = \\left( \\sum _ { w = - \\lfloor k / 2 \\rfloor } ^ { \\lfloor k / 2 \\rfloor } \\sum _ { h = - \\lfloor k / 2 \\rfloor } ^ { \\lfloor k / 2 \\rfloor } \\left. f _ { r \\cdot i + w , r \\cdot j + h , n } \\right. ^ { p } \\right) ^ { 1 / p }\n$$",
+        "text_format": "latex",
+        "bbox": [
+            325,
+            603,
+            673,
+            659
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where the first 2 subscripts index position in the map, and the third subscript indexed the channel. If a convolution operation with the same stride were applied to $\\mathbf { f }$ , we would have a tensor $\\mathbf { c }$ with ",
+        "bbox": [
+            171,
+            664,
+            825,
+            694
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/d41894f353c5f4242f9a88aeebf3cb489c32a246e0ad1a0d0b0a9b8fa2386d78.jpg",
+        "text": "$$\nc _ { i , j , o } = \\sigma \\left( \\sum _ { w = - \\lfloor k / 2 \\rfloor } ^ { \\lfloor k / 2 \\rfloor } \\sum _ { h = - \\lfloor k / 2 \\rfloor } ^ { \\lfloor k / 2 \\rfloor } \\sum _ { n = 1 } ^ { N } t _ { w , h , n , o } f _ { r \\cdot i + w , r \\cdot j + h , n } \\right)\n$$",
+        "text_format": "latex",
+        "bbox": [
+            299,
+            700,
+            696,
+            751
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where $\\mathbf { t }$ is the convolution kernel tensor, and $\\sigma ( \\cdot )$ is an activation function. Thus, a pooling operation can be seen as a convolution operation with uniform kernel tensor and with $L ^ { p }$ -norm as the activation function. ",
+        "bbox": [
+            174,
+            758,
+            825,
+            801
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 EXPERIMENT ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            823,
+            316,
+            838
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 CLASSIFIER ARCHITECTURE AND MULTI-TASK LEARNING ",
+        "text_level": 1,
+        "bbox": [
+            173,
+            854,
+            620,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The architecture of an all-convolution neural network for a single task is shown in Figure 2. The multi-task learning classifier has the same architecture as a single-task classifier except that the width of the output layer is proportional to the number of tasks. The target label is enhanced accordingly by zero-padding. As mentioned before, we combine the training data together, so that the training label dimension is increased from $[ 1 \\times 1 0 ]$ to $[ 1 \\times 2 0 ]$ by zero-padding. This is shown in Figure 3. ",
+        "bbox": [
+            174,
+            881,
+            825,
+            924
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/ac23d7dabb36a6b95a776544ed2d7e602eeee03c2fb47483ca18f15b05732f49.jpg",
+        "image_caption": [
+            "Figure 2: The architecture of an all-convolution neural network for a single task "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            171,
+            103,
+            821,
+            195
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/5f04f6f503b7669c78d2b79422ebfa31a22039d12bda9d76aa4b1e35e970d5b8.jpg",
+        "image_caption": [
+            "Figure 3: The architecture of all-convolution neural network for multi-task learning "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            173,
+            246,
+            821,
+            344
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            169,
+            406,
+            823,
+            435
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Multi-task learning with different combination of datasets are evaluated, namely 3 bi-task learning and 1 tri-task learning. For example, we combine the MNIST train dataset and FashionMNIST train dataset together. Thus, there are 120,000 examples to the bi-task learning network for MNIST and FashionMNIST. Each training example has an input image size of $2 8 \\times 2 8$ and an output 1-hot vector size of 20. ",
+        "bbox": [
+            173,
+            441,
+            825,
+            511
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Each network is trained with 50 epochs. A two-stage learning rate decay scheme is implemented. The initial learning rate is $1 0 ^ { - 3 }$ for the first stage of 25 epochs, and $1 0 ^ { - 5 }$ for the second stage of 25 epochs. When training a model, it is often recommended to lower the learning rate as the training progresses. So, we let learning rate decay in each epochs. The decay rate is $\\frac { \\mathbf { \\bar { \\rho } } _ { 1 } } { 1 + d \\times n }$ where $d$ is the learning rate set in the begin divided by 25. $n$ is the number of current epochs. Figure 4 shows the learning rate scheme. The size of a mini-batch is set to 100, and the Adam optimizer is used. ",
+        "bbox": [
+            173,
+            517,
+            825,
+            603
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "When multi-task learning is complete, the parameters in the network is used to initialize single-task classifiers. Figure 5 illustrates how parameters are handed over. The single-task classifiers are then re-trained to perform their respective classification tasks. ",
+        "bbox": [
+            176,
+            609,
+            823,
+            652
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/b29cbf66c659b1a0765c87255110fee722a9145d3f5c6e61aa9e32ecc4b603c9.jpg",
+        "image_caption": [
+            "Figure 4: The learning rate scheme. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            199,
+            666,
+            797,
+            808
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 RESULTS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            869,
+            279,
+            882
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The experimental results are summarized in Table 3. We are happy to see that multi-task learning works, even though the MNIST, NotMNIST, and FashionMNIST datasets are images from totally different classes. The bi-task learning systems are always better than the single-task systems. Furthermore, the tri-task learning systems are the best, except for the MNIST $( 0 . 0 1 \\%$ difference). The relative reduction in error rates by tri-task learning are respectively $3 1 . 8 \\%$ for MNIST $( 9 9 . 5 6 \\%$ to $9 9 . 7 0 \\%$ ), $1 6 . 4 \\%$ for FashionMNIST $( 9 4 . 3 2 \\%$ to $9 5 . 2 5 \\%$ ), and $8 . 6 \\%$ for NotMNIST $( 9 7 . 2 2 \\%$ to $9 7 . 4 6 \\%$ ). The results confirm that multi-task learning is able to learn representation which is universal and robust to different tasks. ",
+        "bbox": [
+            176,
+            895,
+            823,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/d6df547dae357a04f5bf6418d55fbedf27c8ded7027bb51ff7a15333d33358fe.jpg",
+        "image_caption": [
+            "Figure 5: Passing parameters from multi-task learning network to single-task networks "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            210,
+            101,
+            787,
+            351
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/c56a9f0540db3bde08c24642604d54b2a4381462e08b451d2143ab02cb164068.jpg",
+        "table_caption": [
+            "Table 3: Image recognition accuracy rates for MNIST tasks with multi-task learning "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>dataset</td><td>single-task</td><td>F+N</td><td>M+F</td><td>M+N</td><td>F+N+M</td></tr><tr><td>MNIST</td><td>99.56%</td><td></td><td>99.71%</td><td>99.70%</td><td>99.70%</td></tr><tr><td>NotMNIST</td><td>97.22%</td><td>97.38%</td><td></td><td>97.40%</td><td>97.46%</td></tr><tr><td>FashionMNIST</td><td>94.32%</td><td>95.20%</td><td>95.18%</td><td></td><td>95.25%</td></tr><tr><td></td><td></td><td></td><td></td><td>1</td><td></td></tr></table>",
+        "bbox": [
+            245,
+            428,
+            753,
+            503
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            534,
+            825,
+            617
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To better understand the effect of multi-task learning, we plot the distribution of the high dimensional-data of the classes by the $t$ -distributed stochastic neighbor embedding (t-SNE) (van der Maaten & Hinton (2008)). Figure 6 shows the t-SNE of the output values of several hidden layers of the classifiers. The separation between the learned class manifolds appears to increase with the integration of multi-task learning. The representation learned with multi-task learning in the loop looks better indeed. ",
+        "bbox": [
+            174,
+            623,
+            825,
+            708
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            728,
+            318,
+            744
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we use multi-task learning in pre-training an all-convolution neural network model. We pass the parameters of trained multi-task models to single-task models. Evaluation on MNISTlike datasets show that using multi-task learning can improve image recognition accuracy. The more data we use, the better results we get. This agrees with statistical learning theory that using more data reduces the generalization gap, thus improving test set performance, even if the data comes from a different domain. The classification tasks of the images of digits, letters, and fashion items share parts of their hierarchical representations. By multi-task learning, it is possible to make such common representation robust to help individual classification tasks. ",
+        "bbox": [
+            174,
+            760,
+            825,
+            872
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e0390dfaf853045e89c03ded997c69f36c15fe35724b5b063088a7dcb6a5c9f.jpg",
+        "image_caption": [
+            "Figure 6: Visualization of data manifolds with t-SNE. The left column is the case with multi-task learning, and the right column is the case without multi-task learning. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            264,
+            104,
+            735,
+            491
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            565,
+            285,
+            580
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Yaroslav Bulatov. Notmnist dataset. Available at: http://yaroslavvb. blogspot. it/2011/09/notmnistdataset. html, 2011. ",
+        "bbox": [
+            176,
+            589,
+            823,
+            617
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Rich Caruana. Multitask learning. In Learning to learn, pp. 95–133. Springer, 1998. ",
+        "bbox": [
+            174,
+            628,
+            727,
+            643
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Jia-Ren Chang and Yong-Sheng Chen. Batch-normalized maxout network in network. arXiv preprint arXiv:1511.02583, 2015. ",
+        "bbox": [
+            169,
+            655,
+            823,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Dan Ciregan, Ueli Meier, and Jurgen Schmidhuber. Multi-column deep neural networks for image ¨ classification. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3642–3649, 2012. ",
+        "bbox": [
+            174,
+            694,
+            826,
+            737
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Ronan Collobert and Jason Weston. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proceedings of the 25th International Conference on Machine Learning, pp. 160–167, 2008. ",
+        "bbox": [
+            173,
+            748,
+            826,
+            791
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Li Deng, Geoffrey Hinton, and Brian Kingsbury. New types of deep neural network learning for speech recognition and related applications: An overview. In Proceedings of the IEEE Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8599–8603, 2013. ",
+        "bbox": [
+            174,
+            801,
+            823,
+            845
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to ´ document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. ",
+        "bbox": [
+            173,
+            856,
+            823,
+            885
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Bharath Ramsundar, Steven Kearnes, Patrick Riley, Dale Webster, David Konerding, and Vijay Pande. Massively multitask networks for drug discovery. arXiv preprint arXiv:1502.02072, 2015. ",
+        "bbox": [
+            176,
+            895,
+            823,
+            924
+        ],
+        "page_idx": 5
+    },
+    {
+        "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 Proceedings of the Advances in Neural Information Processing Systems, pp. 91–99, 2015. ",
+        "bbox": [
+            176,
+            103,
+            823,
+            146
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sebastian Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098, 2017. ",
+        "bbox": [
+            171,
+            155,
+            823,
+            184
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ikuro Sato, Hiroki Nishimura, and Kensuke Yokoi. Apac: Augmented pattern classification with neural networks. arXiv preprint arXiv:1505.03229, 2015. ",
+        "bbox": [
+            171,
+            193,
+            823,
+            222
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. ",
+        "bbox": [
+            171,
+            229,
+            823,
+            260
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "L.J.P. van der Maaten and G.E. Hinton. Visualizing high-dimensional data using t-SNE. Machine Learning Research, 9:2579–2605, 2008. ",
+        "bbox": [
+            169,
+            267,
+            823,
+            296
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. Regularization of neural networks using dropconnect. In Proceedings of the 30th international conference on machine learning (ICML), pp. 1058–1066, 2013. ",
+        "bbox": [
+            176,
+            305,
+            823,
+            348
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. ",
+        "bbox": [
+            174,
+            357,
+            821,
+            387
+        ],
+        "page_idx": 6
+    }
+]

parse/train/SJc1hL5ee/SJc1hL5ee.md

@@ -0,0 +1,269 @@
+# FASTTEXT.ZIP: COMPRESSING TEXT CLASSIFICATION MODELS
+
+Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Herve J ´ egou & Tomas Mikolov ´ Facebook AI Research {ajoulin,egrave,bojanowski,matthijs,rvj,tmikolov}@fb.com
+
+# ABSTRACT
+
+We consider the problem of producing compact architectures for text classification, such that the full model fits in a limited amount of memory. After considering different solutions inspired by the hashing literature, we propose a method built upon product quantization to store word embeddings. While the original technique leads to a loss in accuracy, we adapt this method to circumvent quantization artefacts. Combined with simple approaches specifically adapted to text classification, our approach derived from fastText requires, at test time, only a fraction of the memory compared to the original FastText, without noticeably sacrificing quality in terms of classification accuracy. Our experiments carried out on several benchmarks show that our approach typically requires two orders of magnitude less memory than fastText while being only slightly inferior with respect to accuracy. As a result, it outperforms the state of the art by a good margin in terms of the compromise between memory usage and accuracy.
+
+# 1 INTRODUCTION
+
+Text classification is an important problem in Natural Language Processing (NLP). Real world usecases include spam filtering or e-mail categorization. It is a core component in more complex systems such as search and ranking. Recently, deep learning techniques based on neural networks have achieved state of the art results in various NLP applications. One of the main successes of deep learning is due to the effectiveness of recurrent networks for language modeling and their application to speech recognition and machine translation (Mikolov, 2012). However, in other cases including several text classification problems, it has been shown that deep networks do not convincingly beat the prior state of the art techniques (Wang & Manning, 2012; Joulin et al., 2016).
+
+In spite of being (typically) orders of magnitude slower to train than traditional techniques based on n-grams, neural networks are often regarded as a promising alternative due to compact model sizes, in particular for character based models. This is important for applications that need to run on systems with limited memory such as smartphones.
+
+This paper specifically addresses the compromise between classification accuracy and the model size. We extend our previous work implemented in the fastText library1. It is based on n-gram features, dimensionality reduction, and a fast approximation of the softmax classifier (Joulin et al., 2016). We show that a few key ingredients, namely feature pruning, quantization, hashing, and retraining, allow us to produce text classification models with tiny size, often less than 100kB when trained on several popular datasets, without noticeably sacrificing accuracy or speed.
+
+We plan to publish the code and scripts required to reproduce our results as an extension of the fastText library, thereby providing strong reproducible baselines for text classifiers that optimize the compromise between the model size and accuracy. We hope that this will help the engineering community to improve existing applications by using more efficient models.
+
+This paper is organized as follows. Section 2 introduces related work, Section 3 describes our text classification model and explains how we drastically reduce the model size. Section 4 shows the effectiveness of our approach in experiments on multiple text classification benchmarks.
+
+# 2 RELATED WORK
+
+Models for text classification. Text classification is a problem that has its roots in many applications such as web search, information retrieval and document classification (Deerwester et al., 1990; Pang & Lee, 2008). Linear classifiers often obtain state-of-the-art performance while being scalable (Agarwal et al., 2014; Joachims, 1998; Joulin et al., 2016; McCallum & Nigam, 1998). They are particularly interesting when associated with the right features (Wang & Manning, 2012). They usually require storing embeddings for words and n-grams, which makes them memory inefficient.
+
+Compression of language models. Our work is related to compression of statistical language models. Classical approaches include feature pruning based on entropy (Stolcke, 2000) and quantization. Pruning aims to keep only the most important n-grams in the model, leaving out those with probability lower than a specified threshold. Further, the individual n-grams can be compressed by quantizing the probability value, and by storing the n-gram itself more efficiently than as a sequence of characters. Various strategies have been developed, for example using tree structures or hash functions, and are discussed in (Talbot & Brants, 2008).
+
+Compression for similarity estimation and search. There is a large body of literature on how to compress a set of vectors into compact codes, such that the comparison of two codes approximates a target similarity in the original space. The typical use-case of these methods considers an indexed dataset of compressed vectors, and a query for which we want to find the nearest neighbors in the indexed set. One of the most popular is Locality-sensitive hashing (LSH) by Charikar (2002), which is a binarization technique based on random projections that approximates the cosine similarity between two vectors through a monotonous function of the Hamming distance between the two corresponding binary codes. In our paper, LSH refers to this binarization strategy2. Many subsequent works have improved this initial binarization technique, such as spectal hashing (Weiss et al., 2009), or Iterative Quantization (ITQ) (Gong & Lazebnik, 2011), which learns a rotation matrix minimizing the quantization loss of the binarization. We refer the reader to two recent surveys by Wang et al. (2014) and Wang et al. (2015) for an overview of the binary hashing literature.
+
+Beyond these binarization strategies, more general quantization techniques derived from Jegou et al. (2011) offer better trade-offs between memory and the approximation of a distance estimator. The Product Quantization (PQ) method approximates the distances by calculating, in the compressed domain, the distance between their quantized approximations. This method is statistically guaranteed to preserve the Euclidean distance between the vectors within an error bound directly related to the quantization error. The original PQ has been concurrently improved by Ge et al. (2013) and Norouzi & Fleet (2013), who learn an orthogonal transform minimizing the overall quantization loss. In our paper, we will consider the Optimized Product Quantization (OPQ) variant (Ge et al., 2013).
+
+Softmax approximation The aforementioned works approximate either the Euclidean distance or the cosine similarity (both being equivalent in the case of unit-norm vectors). However, in the context of fastText, we are specifically interested in approximating the maximum inner product involved in a softmax layer. Several approaches derived from LSH have been recently proposed to achieve this goal, such as Asymmetric LSH by Shrivastava & Li (2014), subsequently discussed by Neyshabur & Srebro (2015). In our work, since we are not constrained to purely binary codes, we resort a more traditional encoding by employing a magnitude/direction parametrization of our vectors. Therefore we only need to encode/compress an unitary d-dimensional vector, which fits the aforementioned LSH and PQ methods well.
+
+Neural network compression models. Recently, several research efforts have been conducted to compress the parameters of architectures involved in computer vision, namely for state-of-theart Convolutional Neural Networks (CNNs) (Han et al., 2016; Lin et al., 2015). Some use vector quantization (Gong et al., 2014) while others binarize the network (Courbariaux et al., 2016). Denil et al. (2013) show that such classification models are easily compressed because they are overparametrized, which concurs with early observations by LeCun et al. (1990).
+
+Some of these works both aim at reducing the model size and the speed. In our case, since the fastText classifier on which our proposal is built upon is already very efficient, we are primilarly interested in reducing the size of the model while keeping a comparable classification efficiency.
+
+# 3 PROPOSED APPROACH
+
+# 3.1 TEXT CLASSIFICATION
+
+In the context of text classification, linear classifiers (Joulin et al., 2016) remain competitive with more sophisticated, deeper models, and are much faster to train. On top of standard tricks commonly used in linear text classification (Agarwal et al., 2014; Wang & Manning, 2012; Weinberger et al., 2009), Joulin et al. (2016) use a low rank constraint to reduce the computation burden while sharing information between different classes. This is especially useful in the case of a large output space, where rare classes may have only a few training examples. In this paper, we focus on a similar model, that is, which minimizes the softmax loss $\ell$ over $N$ documents:
+
+$$
+\sum _ { n = 1 } ^ { N } \ell ( y _ { n } , B A x _ { n } ) ,
+$$
+
+where $x _ { n }$ is a bag of one-hot vectors and $y _ { n }$ the label of the $n$ -th document. In the case of a large vocabulary and a large output space, the matrices $A$ and $B$ are big and can require gigabytes of memory. Below, we describe how we reduce this memory usage.
+
+# 3.2 BOTTOM-UP PRODUCT QUANTIZATION
+
+Product quantization is a popular method for compressed-domain approximate nearest neighbor search (Jegou et al., 2011). As a compression technique, it approximates a real-valued vector by finding the closest vector in a pre-defined structured set of centroids, referred to as a codebook. This codebook is not enumerated, since it is extremely large. Instead it is implicitly defined by its structure: a $d$ -dimensional vector $x \in \mathbb { R } ^ { d }$ is approximated as
+
+$$
+\hat { x } = \sum _ { i = 1 } ^ { k } q _ { i } ( x ) ,
+$$
+
+where the different subquantizers $q _ { i } : x \mapsto q _ { i } ( x )$ are complementary in the sense that their respective centroids lie in distinct orthogonal subspaces, i.e., $\forall i \neq j$ , $\forall x , y$ , $\langle q _ { i } ( x ) | q _ { j } ( y ) \rangle = 0$ . In the original PQ, the subspaces are aligned with the natural axis, while OPQ learns a rotation, which amounts to alleviating this constraint and to not depend on the original coordinate system. Another way to see this is to consider that PQ splits a given vector $x$ into $k$ subvectors $x ^ { i }$ , $i \stackrel { \cdot } { = } 1 \ldots k$ , each of dimension $d / k$ : $x = [ x ^ { 1 } \dots x ^ { i } \dots x ^ { k } ]$ , and quantizes each sub-vector using a distinct $\mathbf { k }$ -means quantizer. Each subvector $x ^ { i }$ is thus mapped to the closest centroid amongst $2 ^ { b }$ centroids, where $b$ is the number of bits required to store the quantization index of the subquantizer, typically $b = 8$ . The reconstructed vector can take $2 ^ { k b }$ distinct reproduction values, and is stored in $k b$ bits.
+
+PQ estimates the inner product in the compressed domain as
+
+$$
+\boldsymbol { x } ^ { \top } \boldsymbol { y } \approx \hat { \boldsymbol { x } } ^ { \top } \boldsymbol { y } = \sum _ { i = 1 } ^ { k } q _ { i } ( \boldsymbol { x } ^ { i } ) ^ { \top } \boldsymbol { y } ^ { i } .
+$$
+
+This is a straightforward extension of the square L2 distance estimation of Jegou et al. (2011). In practice, the vector estimate $\hat { x }$ is trivially reconstructed from the codes, i.e., from the quantization indexes, by concatenating these centroids.
+
+The two parameters involved in PQ, namely the number of subquantizers $k$ and the number of bits $b$ per quantization index, are typically set to $\dot { k } \in [ 2 , d / 2 ]$ , and $b = 8$ to ensure byte-alignment.
+
+Discussion. PQ offers several interesting properties in our context of text classification. Firstly, the training is very fast because the subquantizers have a small number of centroids, i.e., 256 centroids for $b = 8$ . Secondly, at test time it allows the reconstruction of the vectors with almost no computational and memory overhead. Thirdly, it has been successfully applied in computer vision, offering much better performance than binary codes, which makes it a natural candidate to compress relatively shallow models. As observed by Sanchez & Perronnin (2011), using PQ just before the ´ last layer incurs a very limited loss in accuracy when combined with a support vector machine.
+
+In the context of text classification, the norms of the vectors are widely spread, typically with a ratio of 1000 between the max and the min. Therefore kmeans performs poorly because it optimizes an absolute error objective, so it maps all low-norm vectors to 0. A simple solution is to separate the norm and the angle of the vectors and to quantize them separately. This allows a quantization with no loss of performance, yet requires an extra $b$ bits per vector.
+
+Bottom-up strategy: re-training. The first works aiming at compressing CNN models like the one proposed by (Gong et al., 2014) used the reconstruction from off-the-shelf PQ, i.e., without any re-training. However, as observed in Sablayrolles et al. (2016), when using quantization methods like PQ, it is better to re-train the layers occurring after the quantization, so that the network can re-adjust itself to the quantization. There is a strong argument arguing for this re-training strategy: the square magnitude of vectors is reduced, on average, by the average quantization error for any quantizer satisfying the Lloyd conditions; see Jegou et al. (2011) for details.
+
+This suggests a bottom-up learning strategy where we first quantize the input matrix, then retrain and quantize the output matrix (the input matrix being frozen). Experiments in section 4 show that it is worth adopting this strategy.
+
+Memory savings with PQ. In practice, the bottom-up PQ strategy offers a compression factor of 10 without any noticeable loss of performance. Without re-training, we notice a drop in accuracy between $0 . 1 \%$ and $0 . 5 \%$ , depending on the dataset and setting; see Section 4 and the appendix.
+
+# 3.3 FURTHER TEXT SPECIFIC TRICKS
+
+The memory usage strongly depends on the size of the vocabulary, which can be large in many text classification tasks. While it is clear that a large part of the vocabulary is useless or redundant, directly reducing the vocabulary to the most frequent words is not satisfactory: most of the frequent words, like “the” or “is” are not discriminative, in contrast to some rare words, e.g., in the context of tag prediction. In this section, we discuss a few heuristics to reduce the space taken by the dictionary. They lead to major memory reduction, in extreme cases by a factor 100. We experimentally show that this drastic reduction is complementary with the PQ compression method, meaning that the combination of both strategies reduces the model size by a factor up to $\times 1 0 0 0$ for some datasets.
+
+Pruning the vocabulary. Discovering which word or n-gram must be kept to preserve the overall performance is a feature selection problem. While many approaches have been proposed to select groups of variables during training (Bach et al., 2012; Meier et al., 2008), we are interested in selecting a fixed subset of $K$ words and ngrams from a pre-trained model. This can be achieved by selecting the $K$ embeddings that preserve as much of the model as possible, which can be reduced to selecting the $K$ words and ngrams associated with the highest norms.
+
+While this approach offers major memory savings, it has one drawback occurring in some particular cases: some documents may not contained any of the $K$ best features, leading to a significant drop in performance. It is thus important to keep the $K$ best features under the condition that they cover the whole training set. More formally, the problem is to find a subset $s$ in the feature set $\nu$ that maximizes the sum of their norms $w _ { s }$ under the constraint that all the documents in the training set $\mathcal { D }$ are covered:
+
+$$
+\operatorname* { m a x } _ { S \subseteq \mathcal { V } } \sum _ { s \in \cal { S } } w _ { s } \quad \mathrm { ~ s . t . ~ } \quad | S | \leq K , P 1 _ { \cal { S } } \geq 1 _ { \cal { D } } ,
+$$
+
+where $P$ is a matrix such that $P _ { d s } = 1$ if the $s$ -th feature is in the $d$ -th document, and 0 otherwise. This problem is directly related to set covering problems that are NP-hard (Feige, 1998). Standard greedy approaches require the storing of an inverted index or to do multiple passes over the dataset, which is prohibitive on very large dataset (Chierichetti et al., 2010). This problem can be cast as an instance of online submodular maximization with a rank constraint (Badanidiyuru et al., 2014;
+
+![](images/15d78f88290b96691cc9018373eef6a0532bbee755b0815e5eb2080d3353a831.jpg)  
+Figure 1: Accuracy as a function of the memory per vector/embedding on 3 datasets from Zhang et al. (2015). Note, an extra byte is required when we encode the norm explicitly (”norm”).
+
+Bateni et al., 2010). In our case, we use a simple online parallelizable greedy approach: For each document, we verify if it is already covered by a retained feature and, if not, we add the feature with the highest norm to our set of retained features. If the number of features is below $k$ , we add the features with the highest norm that have not yet been picked.
+
+Hashing trick & Bloom filter. On small models, the dictionary can take a significant portion of the memory. Instead of saving it, we extend the hashing trick used in Joulin et al. (2016) to both words and n-grams. This strategy is also used in Vowpal Wabbit (Agarwal et al., 2014) in the context of online training. This allows us to save around 1-2Mb with almost no overhead at test time (just the cost of computing the hashing function).
+
+Pruning the vocabulary while using the hashing trick requires keeping a list of the indices of the $K$ remaining buckets. At test time, a binary search over the list of indices is required. It has a complexity of $O ( \log ( K ) )$ and a memory overhead of a few hundreds of kilobytes. Using Bloom filters instead reduces the complexity $\mathcal { O } ( 1 )$ at test time and saves a few hundred kilobytes. However, in practice, it degrades performance.
+
+# 4 EXPERIMENTS
+
+This section evaluates the quality of our model compression pipeline and compare it to other compression methods on different text classification problems, and to other compact text classifiers.
+
+Evaluation protocol and datasets. Our experimental pipeline is as follows: we train a model using fastText with the default setting unless specified otherwise. That is 2M buckets, a learning rate of 0.1 and 10 training epochs. The dimensionality $d$ of the embeddings is set to powers of 2 to avoid border effects that could make the interpretation of the results more difficult. As baselines, we use Locality-Sensitive Hashing (LSH) (Charikar, 2002), PQ (Jegou et al., 2011) and OPQ (Ge et al., 2013) (the non-parametric variant). Note that we use an improved version of LSH where random orthogonal matrices are used instead of random matrix projection Jegou et al. (2008). In a first ´ series of experiments, we use the 8 datasets and evaluation protocol of Zhang et al. (2015). These datasets contain few million documents and have at most 10 classes. We also explore the limit of quantization on a dataset with an extremely large output space, that is a tag dataset extracted from the YFCC100M collection (Thomee et al., 2016)3, referred to as FlickrTag in the rest of this paper.
+
+![](images/cdcd3e63a3059e96131d7fa08c0a03f66afb2540d2d14097ce9d24bf7f2608de.jpg)  
+Figure 2: Loss of accuracy as a function of the model size. We compare the compressed model with different level of pruning with NPQ and the full fastText model. We also compare with Zhang et al. (2015) and Xiao & Cho (2016). Note that the size is in log scale.   
+Table 1: Best ranked words w.r.t. entropy (left) and norm (right) on the Amazon full review dataset. We give the rank for both criteria. The norm ranking filters out words carrying little information.
+
+# 4.1 SMALL DATASETS
+
+Compression techniques. We compare three popular methods used for similarity estimation with compact codes: LSH, PQ and OPQ on the datasets released by Zhang et al. (2015). Figure 1 shows the accuracy as a function of the number of bytes used per embedding, which corresponds to the number $k$ of subvectors in the case of PQ and OPQ. See more results in the appendix. As discussed in Section 2, LSH reproduces the cosine similarity and is therefore not adapted to un-normalized data. Therefore we only report results with normalization. Once normalized, PQ and OPQ are almost lossless even when using only $k = 4$ subquantizers per embedding (equivalently, bytes). We observe in practice that using $k = d / 2$ , i.e., half of the components of the embeddings, works well in practice. In the rest of the paper and if not stated otherwise, we focus on this setting. The difference between the normalized versions of PQ and OPQ is limited and depends on the dataset. Therefore we adopt the normalized PQ (NPQ) for the rest of this study, since it is faster to train.
+
+<table><tr><td>word</td><td>Entropy</td><td>Norm</td><td>word</td><td>Entropy</td><td>Norm</td></tr><tr><td>·</td><td>1</td><td>354</td><td>mediocre</td><td>1399</td><td>1</td></tr><tr><td>，</td><td>2</td><td>176</td><td>disappointing</td><td>454</td><td>2</td></tr><tr><td>the</td><td>3</td><td>179</td><td>SO-SO</td><td>2809</td><td>3</td></tr><tr><td>and</td><td>4</td><td>1639</td><td>lacks</td><td>1244</td><td>4</td></tr><tr><td>i</td><td>5</td><td>2374</td><td>worthless</td><td>1757</td><td>5</td></tr><tr><td>a</td><td>6</td><td>970</td><td>dreadful</td><td>4358</td><td>6</td></tr><tr><td>to</td><td>7</td><td>1775</td><td>drm</td><td>6395</td><td>7</td></tr><tr><td>it</td><td>8</td><td>1956</td><td>poorly</td><td>716</td><td>8</td></tr><tr><td>of</td><td>9</td><td>2815</td><td>uninspired</td><td>4245</td><td>9</td></tr><tr><td>this</td><td>10</td><td>3275</td><td>worst</td><td>402</td><td>10</td></tr></table>
+
+Table 2: Performance on very small models. We use a quantization with $k = 1$ , hashing and an extreme pruning. The last row shows the average drop of performance for different size.   
+
+<table><tr><td>Dataset</td><td>full</td><td>64KiB</td><td>32KiB</td><td>16 KiB</td></tr><tr><td>AG</td><td>65M</td><td>92.1 91.4</td><td>90.6</td><td>89.1</td></tr><tr><td>Amazon full</td><td>108M</td><td>60.0 58.8</td><td>56.0</td><td>52.9</td></tr><tr><td>Amazon pol.</td><td>113M</td><td>94.5 93.3</td><td>92.1</td><td>89.3</td></tr><tr><td>DBPedia</td><td>87M</td><td>98.4 98.2</td><td>98.1</td><td>97.4</td></tr><tr><td>Sogou</td><td>73M</td><td>96.4 96.4</td><td>96.3</td><td>95.5</td></tr><tr><td>Yahoo</td><td>122M</td><td>72.1 70.0</td><td>69.0</td><td>69.2</td></tr><tr><td>Yelp full</td><td>78M</td><td>63.8 63.2</td><td>62.4</td><td>58.7</td></tr><tr><td>Yelp pol.</td><td>77M</td><td>95.7 95.3</td><td>94.9</td><td>93.2</td></tr><tr><td>Average diff. [%]</td><td></td><td>0</td><td>-0.8 -1.7</td><td>-3.5</td></tr></table>
+
+Pruning. Figure 2 shows the performance of our model with different sizes. We fix $k = d / 2$ and use different pruning thresholds. NPQ offers a compression rate of $\times 1 0$ compared to the full model. As the pruning becomes more agressive, the overall compression can increase up up to $\times 1 , 0 0 0$ with little drop of performance and no additional overhead at test time. In fact, using a smaller dictionary makes the model faster at test time. We also compare with character-level Convolutional Neural Networks (CNN) (Zhang et al., 2015; Xiao & Cho, 2016). They are attractive models for text classification because they achieve similar performance with less memory usage than linear models (Xiao & Cho, 2016). Even though fastText with the default setting uses more memory, NPQ is already on par with CNNs’ memory usage. Note that CNNs are not quantized, and it would be worth seeing how much they can be quantized with no drop of performance. Such a study is beyond the scope of this paper. Our pruning is based on the norm of the embeddings according to the guidelines of Section 3.3. Table 1 compares the ranking obtained with norms to the ranking obtained using entropy, which is commonly used in unsupervised settings Stolcke (2000).
+
+Extreme compression. Finally, in Table 2, we explore the limit of quantized model by looking at the performance obtained for models under 64KiB. Surprisingly, even at 64KiB and 32KiB, the drop of performance is only around $0 . 8 \%$ and $1 . 7 \%$ despite a compression rate of $\times 1 , 0 0 0 - 4 , 0 0 0$ .
+
+# 4.2 LARGE DATASET: FLICKRTAG
+
+In this section, we explore the limit of compression algorithms on very large datasets. Similar to Joulin et al. (2016), we consider a hashtag prediction dataset containing 312, 116 labels. We set the minimum count for words at 10, leading to a dictionary of 1, 427, 667 words. We take 10M buckets for n-grams and a hierarchical softmax. We refer to this dataset as FlickrTag.
+
+Output encoding. We are interested in understanding how the performance degrades if the classifier is also quantized (i.e., the matrix $B$ in Eq. 1) and when the pruning is at the limit of the minimum number of features required to cover the full dataset.   
+
+<table><tr><td>Model</td><td>k</td><td>norm</td><td>retrain</td><td>Acc.</td><td>Size</td></tr><tr><td>full (uncompressed)</td><td></td><td></td><td></td><td>45.4</td><td>12 GiB</td></tr><tr><td>Input</td><td>128</td><td></td><td></td><td>45.0</td><td>1.7 GiB</td></tr><tr><td>Input</td><td>128</td><td>X</td><td></td><td>45.3</td><td>1.8 GiB</td></tr><tr><td>Input</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>1.8 GiB</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td></td><td>45.2</td><td>1.5 GiB</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td>X</td><td>45.4</td><td>1.5 GiB</td></tr></table>
+
+Table 3: FlickrTag: Influence of quantizing the output matrix on performance. We use PQ for quantization with an optional normalization. We also retrain the output matrix after quantizing the input one. The ”norm” refers to the separate encoding of the magnitude and angle, while ”retrain” refers to the re-training bottom-up PQ method described in Section 3.2.
+
+Table 3 shows that quantizing both the “input” matrix (i.e., $A$ in Eq. 1) and the “output” matrix (i.e., $B$ ) does not degrade the performance compared to the full model. We use embeddings with $d = 2 5 6$ dimensions and use $k = d / 2$ subquantizers. We do not use any text specific tricks, which leads to a compression factor of 8. Note that even if the output matrix is not retrained over the embeddings, the performance is only $0 . 2 \%$ away from the full model. As shown in the Appendix, using less subquantizers significantly decreases the performance for a small memory gain.
+
+<table><tr><td>Model</td><td>full</td><td colspan="2"> Entropy pruning</td><td colspan="2">Norm pruning</td><td colspan="2">Max-Cover pruning</td></tr><tr><td>#embeddings Memory</td><td>11.5M 12GiB</td><td>2M 297MiB</td><td>1M 174MiB</td><td>2M 305MiB</td><td>1M 179MiB</td><td>2M 305MiB</td><td>1M 179MiB</td></tr><tr><td>Coverage [%]</td><td>88.4</td><td>70.5</td><td>70.5</td><td>73.2</td><td>61.9</td><td>88.4</td><td>88.4</td></tr><tr><td>Accuracy</td><td>45.4</td><td>32.1</td><td>30.5</td><td>41.6</td><td>35.8</td><td>45.5</td><td>43.9</td></tr></table>
+
+Table 4: FlickrTag: Comparison of entropy pruning, norm pruning and max-cover pruning methods.   
+We show the coverage of the test set for each method.
+
+Pruning. Table 4 shows how the performance evolves with pruning. We measure this effect on top of a fully quantized model. The full model misses $1 1 . 6 \%$ of the test set because of missing words (some documents are either only composed of hashtags or have only rare words). There are 312, 116 labels and thus it seems reasonable to keep embeddings in the order of the million. A naive pruning with 1M features misses about $3 0 - 4 0 \%$ of the test set, leading to a significant drop of performance. On the other hand, even though the max-coverage pruning approach was set on the train set, it does not suffer from any coverage loss on the test set. This leads to a smaller drop of performance. If the pruning is too aggressive, however, the coverage decreases significantly.
+
+# 5 FUTURE WORK
+
+It may be possible to obtain further reduction of the model size in the future. One idea is to condition the size of the vectors (both for the input features and the labels) based on their frequency (Chen et al., 2015; Grave et al., 2016). For example, it is probably not worth representing the rare labels by full 256-dimensional vectors in the case of the FlickrTag dataset. Thus, conditioning the vector size on the frequency and norm seems like an interesting direction to explore in the future.
+
+We may also consider combining the entropy and norm pruning criteria: instead of keeping the features in the model based just on the frequency or the norm, we can use both to keep a good set of features. This could help to keep features that are both frequent and discriminative, and thereby to reduce the coverage problem that we have observed.
+
+Additionally, instead of pruning out the less useful features, we can decompose them into smaller units (Mikolov et al., 2012). For example, this can be achieved by splitting every non-discriminative word into a sequence of character trigrams. This could help in cases where training and test examples are very short (for example just a single word).
+
+# 6 CONCLUSION
+
+In this paper, we have presented several simple techniques to reduce, by several orders of magnitude, the memory complexity of certain text classifiers without sacrificing accuracy nor speed. This is achieved by applying discriminative pruning which aims to keep only important features in the trained model, and by performing quantization of the weight matrices and hashing of the dictionary.
+
+We will publish the code as an extension of the fastText library. We hope that our work will serve as a baseline to the research community, where there is an increasing interest for comparing the performance of various deep learning text classifiers for a given number of parameters. Overall, compared to recent work based on convolutional neural networks, fastText.zip is often more accurate, while requiring several orders of magnitude less time to train on common CPUs, and incurring a fraction of the memory complexity.
+
+# REFERENCES
+
+Alekh Agarwal, Olivier Chapelle, Miroslav Dud´ık, and John Langford. A reliable effective terascale linear learning system. Journal of Machine Learning Research, 15(1):1111–1133, 2014.
+
+Francis Bach, Rodolphe Jenatton, Julien Mairal, and Guillaume Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends $\textsuperscript { \textregistered }$ in Machine Learning, 4(1):1–106, 2012.
+
+Ashwinkumar Badanidiyuru, Baharan Mirzasoleiman, Amin Karbasi, and Andreas Krause. Streaming submodular maximization: Massive data summarization on the fly. In SIGKDD, pp. 671–680. ACM, 2014.
+
+Mohammad Hossein Bateni, Mohammad Taghi Hajiaghayi, and Morteza Zadimoghaddam. Submodular secretary problem and extensions. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pp. 39–52. Springer, 2010.
+
+Moses S. Charikar. Similarity estimation techniques from rounding algorithms. In STOC, pp. 380– 388, May 2002.
+
+Welin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural language models. arXiv preprint arXiv:1512.04906, 2015.
+
+Flavio Chierichetti, Ravi Kumar, and Andrew Tomkins. Max-cover in map-reduce. In International Conference on World Wide Web, 2010.
+
+Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks: Training neural networks with weights and activations constrained to $+ 1$ or $^ { - 1 }$ . arXiv preprint arXiv:1602.02830, 2016.
+
+M. Datar, N. Immorlica, P. Indyk, and V.S. Mirrokni. Locality-sensitive hashing scheme based on pstable distributions. In Proceedings of the Symposium on Computational Geometry, pp. 253–262, 2004.
+
+Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. Indexing by latent semantic analysis. Journal of the American society for information science, 1990.
+
+Misha Denil, Babak Shakibi, Laurent Dinh, Marc-Aurelio Ranzato, and Nando et all de Freitas. Predicting parameters in deep learning. In NIPS, pp. 2148–2156, 2013.
+
+Uriel Feige. A threshold of ln n for approximating set cover. JACM, 45(4):634–652, 1998.
+
+Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. Optimized product quantization for approximate nearest neighbor search. In CVPR, June 2013.
+
+Yunchao Gong and Svetlana Lazebnik. Iterative quantization: A procrustean approach to learning binary codes. In CVPR, June 2011.
+
+Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014.
+
+Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´ softmax approximation for gpus. arXiv preprint arXiv:1609.04309, 2016.
+
+Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016.
+
+Herve J ´ egou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric ´ consistency for large scale image search. In ECCV, October 2008.
+
+Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor ´ search. IEEE Trans. PAMI, January 2011.
+
+Thorsten Joachims. Text categorization with support vector machines: Learning with many relevant features. Springer, 1998.
+
+Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016.
+
+Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. NIPS, 2:598–605, 1990.
+
+Zhouhan Lin, Matthieu Courbariaux, Roland Memisevic, and Yoshua Bengio. Neural networks with few multiplications. arXiv preprint arXiv:1510.03009, 2015.
+
+Andrew McCallum and Kamal Nigam. A comparison of event models for naive bayes text classification. In AAAI workshop on learning for text categorization, 1998.
+
+Lukas Meier, Sara Van De Geer, and Peter Buhlmann. The group lasso for logistic regression. ¨ Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70(1):53–71, 2008.
+
+Tomas Mikolov. Statistical language models based on neural networks. In PhD thesis. VUT Brno, 2012.
+
+Tomas Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and J Cernocky. Subword language modeling with neural networks. preprint, 2012.
+
+Behnam Neyshabur and Nathan Srebro. On symmetric and asymmetric lshs for inner product search. In ICML, pp. 1926–1934, 2015.
+
+Mohammad Norouzi and David Fleet. Cartesian k-means. In CVPR, June 2013.
+
+Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations and trends in information retrieval, 2008.
+
+Alexandre Sablayrolles, Matthijs Douze, Herve J ´ egou, and Nicolas Usunier. How should we evalu-´ ate supervised hashing? arXiv preprint arXiv:1609.06753, 2016.
+
+Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale im- ´ age classification. In CVPR, 2011.
+
+Anshumali Shrivastava and Ping Li. Asymmetric LSH for sublinear time maximum inner product search. In NIPS, pp. 2321–2329, 2014.
+
+Andreas Stolcke. Entropy-based pruning of backoff language models. arXiv preprint cs/0006025, 2000.
+
+David Talbot and Thorsten Brants. Randomized language models via perfect hash functions. In ACL, 2008.
+
+Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In Communications of the ACM, 2016.
+
+Jingdong Wang, Heng Tao Shen, Jingkuan Song, and Jianqiu Ji. Hashing for similarity search: A survey. arXiv preprint arXiv:1408.2927, 2014.
+
+Jun Wang, Wei Liu, Sanjiv Kumar, and Shih-Fu Chang. Learning to hash for indexing big data - A survey. CoRR, abs/1509.05472, 2015.
+
+Sida Wang and Christopher D Manning. Baselines and bigrams: Simple, good sentiment and topic classification. In ACL, 2012.
+
+Kilian Q Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. Feature hashing for large scale multitask learning. In ICML, 2009.
+
+Yair Weiss, Antonio Torralba, and Rob Fergus. Spectral hashing. In NIPS, December 2009.
+
+Yijun Xiao and Kyunghyun Cho. Efficient character-level document classification by combining convolution and recurrent layers. arXiv preprint arXiv:1602.00367, 2016.
+
+Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In NIPS, 2015.
+
+# APPENDIX
+
+In the appendix, we show some additional results. The model used in these experiments only had 1M ngram buckets. In Table 5, we show a thorough comparison of LSH, PQ and OPQ on 8 different datasets. Table 7 summarizes the comparison with CNNs in terms of accuracy and size. Table 8 show a thorough comparison of the hashing trick and the Bloom filters.   
+
+<table><tr><td>Quant. k</td><td>ULIOU</td><td></td><td>AG</td><td></td><td>Amz. f.</td><td></td><td>Amz. p.</td><td></td><td>DBP</td><td></td><td></td><td>Sogou</td><td></td><td>Yah.</td><td></td><td>Yelp f.</td><td></td><td>Yelp p.</td></tr><tr><td>full full,nodict</td><td></td><td></td><td>92.1 36M 92.1 34M</td><td></td><td>59.897M 59.9</td><td>78M</td><td>94.5 94.5</td><td>104M 83M</td><td>98.467M 98.4</td><td>56M</td><td>96.347M 96.3</td><td>42M</td><td>72 72.2</td><td>120M 91M</td><td>63.756M 63.648M</td><td></td><td>95.7 53M 95.646M</td><td></td></tr><tr><td>LSH</td><td>8</td><td></td><td>88.7</td><td>8.5M</td><td>51.3</td><td>20M</td><td>90.3</td><td>21M</td><td>92.7</td><td>14M</td><td>94.2</td><td>11M</td><td>54.8</td><td>23M</td><td>56.7</td><td>12M</td><td>92.2</td><td>12M</td></tr><tr><td>PQ</td><td>8</td><td></td><td>91.7</td><td>8.5M</td><td>59.3</td><td>20M</td><td>94.4</td><td>21M</td><td>97.4</td><td>14M</td><td>96.1</td><td>11M</td><td>71.3</td><td>23M</td><td>62.8</td><td>12M</td><td>95.4</td><td>12M</td></tr><tr><td>OPQ</td><td>8</td><td></td><td>91.9</td><td>8.5M</td><td>59.3</td><td>20M</td><td>94.4</td><td>21M</td><td>96.9</td><td>14M</td><td>95.8</td><td>11M</td><td>71.4</td><td>23M</td><td>62.5</td><td>12M</td><td>95.4</td><td>12M</td></tr><tr><td>LSH</td><td>8</td><td>X</td><td>91.9</td><td>9.5M</td><td>59.4</td><td>22M</td><td>94.5</td><td>24M</td><td>97.8</td><td>16M</td><td>96.2</td><td>12M</td><td>71.6</td><td>26M</td><td>63.4</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>PQ</td><td>8</td><td>X</td><td>92.0</td><td>9.5M</td><td>59.8</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.1</td><td>26M</td><td>63.7</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>OPQ</td><td>8</td><td>X</td><td>92.1</td><td>9.5M</td><td>59.9</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.2</td><td>26M</td><td>63.6</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>LSH</td><td>4</td><td></td><td>88.3</td><td>4.3M</td><td>50.5</td><td>9.7M</td><td>88.9</td><td>11M</td><td>91.6</td><td>7.0M</td><td>94.3</td><td>5.3M</td><td>54.6</td><td>12M</td><td>56.5</td><td>6.0M</td><td>92.9</td><td>5.7M</td></tr><tr><td>PQ</td><td>4</td><td></td><td>91.6</td><td>4.3M</td><td>59.2</td><td>9.7M</td><td>94.4</td><td>11M</td><td>96.3</td><td>7.0M</td><td>96.1</td><td>5.3M</td><td>71.0</td><td>12M</td><td>62.2</td><td>6.0M</td><td>95.4</td><td>5.7M</td></tr><tr><td>OPQ</td><td>4</td><td></td><td>91.7</td><td>4.3M</td><td>59.0</td><td>9.7M</td><td>94.4</td><td>11M</td><td>96.9</td><td>7.0M</td><td>95.6</td><td>5.3M</td><td>71.2</td><td>12M</td><td>62.6</td><td>6.0M</td><td>95.4</td><td>5.7M</td></tr><tr><td>LSH</td><td>4</td><td>X</td><td>92.1</td><td>5.3M</td><td>59.2</td><td>13M</td><td>94.4</td><td>13M</td><td>97.7</td><td>8.8M</td><td>96.2</td><td>6.6M</td><td>71.1</td><td>15M</td><td>63.1</td><td>7.4M</td><td>95.5</td><td>7.2M</td></tr><tr><td>PQ</td><td>4</td><td>X</td><td>92.1</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.4</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72.0</td><td>15M</td><td>63.6</td><td>7.5M</td><td></td><td>95.67.2M</td></tr><tr><td>OPQ</td><td>4</td><td>X</td><td>92.2</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.3</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72.1</td><td>15M</td><td>63.7</td><td>7.5M</td><td>95.6</td><td>7.2M</td></tr><tr><td>LSH</td><td>2</td><td></td><td>87.7</td><td>2.2M</td><td>50.1</td><td>4.9M</td><td>88.9</td><td>5.2M</td><td>90.6</td><td>3.5M</td><td>93.9</td><td>2.7M</td><td>51.4</td><td>5.7M</td><td>56.6</td><td>3.0M</td><td>91.3</td><td>2.9M</td></tr><tr><td>PQ</td><td>2</td><td></td><td>91.1</td><td>2.2M</td><td>58.7</td><td>4.9M</td><td>94.4</td><td>5.2M</td><td>87.1</td><td>3.6M</td><td>95.3</td><td>2.7M</td><td>69.5</td><td>5.7M</td><td>62.1</td><td>3.0M</td><td>95.4</td><td>2.9M</td></tr><tr><td>OPQ</td><td>2</td><td></td><td>91.4</td><td>2.2M</td><td>58.2</td><td>4.9M</td><td>94.3</td><td>5.2M</td><td>91.6</td><td>3.6M</td><td>94.2</td><td>2.7M</td><td>69.6</td><td>5.7M</td><td>62.1</td><td>3.0M</td><td>95.4</td><td>2.9M</td></tr><tr><td>LSH</td><td>2</td><td>X</td><td>91.8</td><td>3.2M</td><td>58.6</td><td>7.3M</td><td>94.3</td><td>7.8M</td><td>97.1</td><td>5.3M</td><td>96.1</td><td>4.0M</td><td>69.7</td><td>8.6M</td><td>62.7</td><td>4.5M</td><td></td><td>95.54.3M</td></tr><tr><td>PQ</td><td>2</td><td>X</td><td>91.9</td><td>3.2M</td><td>59.6</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.3</td><td>4.0M</td><td>71.3</td><td>8.6M</td><td>63.44.5M</td><td></td><td></td><td>95.64.3M</td></tr><tr><td>OPQ</td><td>2</td><td>X</td><td>92.1</td><td>3.2M</td><td>59.5</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.2</td><td>4.0M</td><td>71.5</td><td>8.6M</td><td>63.44.5M</td><td></td><td></td><td>95.64.3M</td></tr></table>
+
+Table 5: Comparison between standard quantization methods. The original model has a dimensionality of 8 and 2M buckets. Note that all of the methods are without dictionary.   
+
+<table><tr><td>k</td><td>co</td><td colspan="2">AG</td><td colspan="2">Amz. f.</td><td colspan="2">Amz. p.</td><td colspan="2">DBP</td><td colspan="2">Sogou</td><td colspan="2">Yah.</td><td colspan="2">Yelp f.</td><td colspan="2">Yelp p.</td></tr><tr><td></td><td>full, nodict</td><td>92.1</td><td>34M</td><td>59.8</td><td>78M</td><td>94.5</td><td>83M</td><td>98.4</td><td>56M</td><td>96.3</td><td>42M</td><td>72.2</td><td>91M</td><td>63.7</td><td>48M</td><td>95.6</td><td>46M</td></tr><tr><td>8</td><td>full</td><td>92.0</td><td>9.5M</td><td>59.8</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.1</td><td>26M</td><td>63.7</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>4</td><td>full</td><td>92.1</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.4</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72</td><td>15M</td><td>63.6</td><td>7.5M</td><td>95.6</td><td>7.2M</td></tr><tr><td>2</td><td>full</td><td>91.9</td><td>3.2M</td><td>59.6</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.3</td><td>4.0M</td><td>71.3</td><td>8.6M</td><td>63.4</td><td>4.5M</td><td>95.6</td><td>4.3M</td></tr><tr><td>8</td><td>200K</td><td>92.0</td><td>2.5M</td><td>59.7</td><td>2.5M</td><td>94.3</td><td>2.5M</td><td>98.5</td><td>2.5M</td><td>96.6</td><td>2.5M</td><td>71.8</td><td>2.5M</td><td>63.3</td><td>2.5M</td><td>95.6</td><td>2.5M</td></tr><tr><td>8</td><td>100K</td><td>91.9</td><td>1.3M</td><td>59.5</td><td>1.3M</td><td>94.3</td><td>1.3M</td><td>98.5</td><td>1.3M</td><td>96.6</td><td>1.3M</td><td>71.6</td><td>1.3M</td><td>63.4</td><td>1.3M</td><td>95.6</td><td>1.3M</td></tr><tr><td>8</td><td>50K</td><td>91.7</td><td>645K</td><td>59.7</td><td>645K</td><td>94.3</td><td>644K</td><td>98.5</td><td>645K</td><td>96.6</td><td>645K</td><td>71.5</td><td>645K</td><td>63.2</td><td>645K</td><td>95.6</td><td>644K</td></tr><tr><td>8</td><td>10K</td><td>91.3</td><td>137K</td><td>58.6</td><td>137K</td><td>93.2</td><td>137K</td><td>98.5</td><td>137K</td><td>96.5</td><td>137K</td><td>71.3</td><td>137K</td><td>63.3</td><td>137K</td><td>95.4</td><td>137K</td></tr><tr><td>4</td><td>200K</td><td>92.0</td><td>1.8M</td><td>59.7</td><td>1.8M</td><td>94.3</td><td>1.8M</td><td>98.5</td><td>1.8M</td><td>96.6</td><td>1.8M</td><td>71.7</td><td>1.8M</td><td>63.3</td><td>1.8M</td><td>95.6</td><td>1.8M</td></tr><tr><td>4</td><td>100K</td><td>91.9</td><td>889K</td><td>59.5</td><td>889K</td><td>94.4</td><td>889K</td><td>98.5</td><td>889K</td><td>96.6</td><td>889K</td><td>71.7</td><td>889K</td><td>63.4</td><td>889K</td><td>95.6</td><td>889K</td></tr><tr><td>4</td><td>50K</td><td>91.7</td><td>449K</td><td>59.6</td><td>449K</td><td>94.3</td><td>449K</td><td>98.5</td><td>450K</td><td>96.6</td><td>449K</td><td>71.4</td><td>450K</td><td>63.2</td><td>449K</td><td>95.5</td><td>449K</td></tr><tr><td>4</td><td>10K</td><td>91.5</td><td>98K</td><td>58.6</td><td>98K</td><td>93.2</td><td>98K</td><td>98.5</td><td>98K</td><td>96.5</td><td>98K</td><td>71.2</td><td>98K</td><td>63.3</td><td>98K</td><td>95.4</td><td>98K</td></tr><tr><td></td><td>200K</td><td>91.9</td><td>1.4M</td><td>59.6</td><td>1.4M</td><td>94.3</td><td>1.4M</td><td>98.4</td><td>1.4M</td><td>96.5</td><td>1.4M</td><td>71.5</td><td>1.4M</td><td>63.2</td><td>1.4M</td><td>95.5</td><td>1.4M</td></tr><tr><td></td><td>100K</td><td>91.6</td><td>693K</td><td>59.5</td><td>693K</td><td>94.3</td><td>693K</td><td>98.4</td><td>694K</td><td>96.6</td><td>693K</td><td>71.1</td><td>694K</td><td>63.2</td><td>693K</td><td>95.6</td><td>693K</td></tr><tr><td>2222</td><td>50K</td><td>91.6</td><td>352K</td><td>59.6</td><td>352K</td><td>94.3</td><td>352K</td><td>98.4</td><td>352K</td><td>96.5</td><td>352K</td><td>71.1</td><td>352K</td><td>63.2</td><td>352K</td><td>95.6</td><td>352K</td></tr><tr><td></td><td>10K</td><td>91.3</td><td>78K</td><td>58.5</td><td>78K</td><td>93.2</td><td>78K</td><td>98.4</td><td>79K</td><td>96.5</td><td>78K</td><td>70.8</td><td>78K</td><td>63.2</td><td>78K</td><td>95.3</td><td>78K</td></tr></table>
+
+Table 6: Comparison with different quantization and level of pruning. “co” is the cut-off parameter of the pruning.
+
+<table><tr><td>Dataset</td><td>Zhang et al. (2015)</td><td></td><td>Xiao &amp; Cho (2016)</td><td></td><td>fastText+PQ,k= d/2</td></tr><tr><td>AG</td><td>90.2</td><td>108M</td><td>91.4</td><td>80M</td><td>91.9</td></tr><tr><td>Amz. f.</td><td>59.5</td><td>10.8M</td><td>59.2</td><td>1.6M</td><td>59.6 449K</td></tr><tr><td>Amz. p.</td><td>94.5</td><td>10.8M</td><td>94.1</td><td>1.6M</td><td>449K</td></tr><tr><td>DBP</td><td>98.3</td><td>108M</td><td>98.6</td><td>1.2M</td><td>98K</td></tr><tr><td>Sogou</td><td>95.1</td><td>108M</td><td>95.2</td><td>1.6M</td><td>98K</td></tr><tr><td>Yah.</td><td>70.5</td><td>108M</td><td>71.4</td><td>80M</td><td>889K</td></tr><tr><td>Yelp f.</td><td>61.6</td><td>108M</td><td>61.8</td><td>1.4M</td><td>98K</td></tr><tr><td>Yelp p.</td><td>94.8</td><td>108M</td><td>94.5</td><td>1.2M</td><td>449K</td></tr></table>
+
+Table 7: Comparison between CNNs and fastText with and without quantization. The numbers for Zhang et al. (2015) are reported from Xiao & Cho (2016). Note that for the CNNs, we report the size of the model under the assumption that they use float32 storage. For fastText $( + \mathrm { P Q } )$ we report the memory used in RAM at test time.
+
+Table 8: Comparison with and without Bloom filters. For NPQ, we set $d = 8$ and $k = 2$ .   
+
+<table><tr><td>Quant.</td><td>1ooom</td><td>co</td><td colspan="2">AG</td><td colspan="2">Amz. f.</td><td colspan="2">Amz. p.</td><td colspan="2">DBP</td><td colspan="2">Sogou</td><td colspan="2">Yah.</td><td colspan="2">Yelp f.</td><td colspan="2">Yelp p.</td></tr><tr><td>full,nodict</td><td></td><td></td><td>92.1</td><td>34M</td><td>59.8</td><td>78M</td><td>94.5</td><td>83M</td><td>98.4</td><td>56M</td><td>96.3</td><td>42M</td><td>72.2</td><td>91M</td><td>63.7</td><td>48M</td><td>95.6</td><td>46M</td></tr><tr><td>NPQ</td><td></td><td>200K</td><td>91.9</td><td>1.4M</td><td>59.6</td><td>1.4M</td><td>94.3</td><td>1.4M</td><td>98.4</td><td>1.4M</td><td>96.5</td><td>1.4M</td><td>71.5</td><td>1.4M</td><td>63.2</td><td>1.4M</td><td>95.5</td><td>1.4M</td></tr><tr><td>NPQ</td><td>X</td><td>200K</td><td>92.2</td><td>830K</td><td>59.3</td><td>830K</td><td>94.1</td><td>830K</td><td>98.4</td><td>830K</td><td>96.5</td><td>830K</td><td>70.7</td><td>830K</td><td>63.0</td><td>830K</td><td>95.5</td><td>830K</td></tr><tr><td>NPQ</td><td></td><td>100K</td><td>91.6</td><td>693K</td><td>59.5</td><td>693K</td><td>94.3</td><td>693K</td><td>98.4</td><td>694K</td><td>96.6</td><td>693K</td><td>71.1</td><td>694K</td><td>63.2</td><td>693K</td><td>95.6</td><td>693K</td></tr><tr><td>NPQ</td><td>X</td><td>100K</td><td>91.8</td><td>420K</td><td>59.1</td><td>420K</td><td>93.9</td><td>420K</td><td>98.4</td><td>420K</td><td>96.5</td><td>420K</td><td>70.6</td><td>420K</td><td>62.8</td><td>420K</td><td>95.3</td><td>420K</td></tr><tr><td>NPQ</td><td></td><td>50K</td><td>91.6</td><td>352K</td><td>59.6</td><td>352K</td><td>94.3</td><td>352K</td><td>98.4</td><td>352K</td><td>96.5</td><td>352K</td><td>71.1</td><td>352K</td><td>63.2</td><td>352K</td><td>95.6</td><td>352K</td></tr><tr><td>NPQ</td><td>X</td><td>50K</td><td>91.5</td><td>215K</td><td>58.8</td><td>215K</td><td>93.6</td><td>215K</td><td>98.3</td><td>215K</td><td>96.5</td><td>215K</td><td>70.1</td><td>215K</td><td>62.7</td><td>215K</td><td>95.1</td><td>215K</td></tr><tr><td>NPQ</td><td></td><td>10K</td><td>91.3</td><td>78K</td><td>58.5</td><td>78K</td><td>93.2</td><td>78K</td><td>98.4</td><td>79K</td><td>96.5</td><td>78K</td><td>70.8</td><td>78K</td><td>63.2</td><td>78K</td><td>95.3</td><td>78K</td></tr><tr><td>NPQ</td><td>X</td><td>10K</td><td>90.8</td><td>51K</td><td>56.8</td><td>51K</td><td>91.7</td><td>51K</td><td>98.1</td><td>51K</td><td>96.1</td><td>51K</td><td>68.7</td><td>51K</td><td>61.7</td><td>51K</td><td>94.5</td><td>51K</td></tr></table>
+
+<table><tr><td>Model</td><td>k</td><td>norm</td><td>retrain</td><td>Acc.</td><td>Size</td></tr><tr><td>full</td><td></td><td></td><td></td><td>45.4</td><td>12G</td></tr><tr><td>Input</td><td>128</td><td></td><td></td><td>45.0</td><td>1.7G</td></tr><tr><td>Input</td><td>128</td><td>X</td><td></td><td>45.3</td><td>1.8G</td></tr><tr><td>Input</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>1.8G</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td></td><td>45.2</td><td>1.5G</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td>X</td><td>45.4</td><td>1.5G</td></tr><tr><td>Input+Output, co=2M</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>305M</td></tr><tr><td>Input+Output, n co=1M</td><td>128</td><td>X</td><td>X</td><td>43.9</td><td>179M</td></tr><tr><td>Input</td><td>64</td><td></td><td></td><td>44.0</td><td>1.1G</td></tr><tr><td>Input</td><td>64</td><td>X</td><td></td><td>44.7</td><td>1.1G</td></tr><tr><td>Input</td><td>64</td><td>X</td><td></td><td>44.9</td><td>1.1G</td></tr><tr><td>Input+Output</td><td>64</td><td>X</td><td></td><td>44.6</td><td>784M</td></tr><tr><td>Input+Output</td><td>64</td><td>X</td><td>X</td><td>44.8</td><td>784M</td></tr><tr><td>Input+Output, co=2M</td><td>64</td><td>X</td><td></td><td>42.5</td><td>183M</td></tr><tr><td>Input+Output, co=1M</td><td>64</td><td>X</td><td></td><td>39.9</td><td>118M</td></tr><tr><td>Input+Output, co=2M</td><td>64</td><td>X</td><td>X</td><td>45.0</td><td>183M</td></tr><tr><td>Input+Output,co=1M</td><td>64</td><td>X</td><td>X</td><td>43.4</td><td>118M</td></tr><tr><td>Input</td><td>32</td><td></td><td></td><td>40.5</td><td>690M</td></tr><tr><td>Input</td><td>32</td><td>X</td><td></td><td>42.4</td><td>701M</td></tr><tr><td>Input</td><td>32</td><td>X</td><td>X</td><td>42.9</td><td>701M</td></tr><tr><td>Input+Output</td><td>32</td><td>X</td><td></td><td>42.3</td><td>435M</td></tr><tr><td>Input+Output</td><td>32</td><td>X</td><td>X</td><td>42.8</td><td>435M</td></tr><tr><td>Input+Output,co=2M</td><td>32</td><td>X</td><td></td><td>35.0</td><td>122M</td></tr><tr><td>Input+Output, co=1M</td><td>32</td><td>X</td><td></td><td>32.6</td><td>88M</td></tr><tr><td>Input+Output, co=2M</td><td>32</td><td>X</td><td>X</td><td>43.3</td><td>122M</td></tr><tr><td>Input+Output, co=1M</td><td>32</td><td>X</td><td>X</td><td>41.6</td><td>88M</td></tr></table>
+
+Table 9: FlickrTag: Comparison for a large dataset of (i) different quantization methods and parameters, (ii) with or without re-training.

parse/train/SJc1hL5ee/SJc1hL5ee_content_list.json

@@ -0,0 +1,1428 @@
+[
+    {
+        "type": "text",
+        "text": "FASTTEXT.ZIP: COMPRESSING TEXT CLASSIFICATION MODELS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            99,
+            735,
+            146
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Herve J ´ egou & Tomas Mikolov ´ Facebook AI Research {ajoulin,egrave,bojanowski,matthijs,rvj,tmikolov}@fb.com ",
+        "bbox": [
+            183,
+            169,
+            872,
+            213
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ABSTRACT ",
+        "text_level": 1,
+        "bbox": [
+            454,
+            248,
+            544,
+            263
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We consider the problem of producing compact architectures for text classification, such that the full model fits in a limited amount of memory. After considering different solutions inspired by the hashing literature, we propose a method built upon product quantization to store word embeddings. While the original technique leads to a loss in accuracy, we adapt this method to circumvent quantization artefacts. Combined with simple approaches specifically adapted to text classification, our approach derived from fastText requires, at test time, only a fraction of the memory compared to the original FastText, without noticeably sacrificing quality in terms of classification accuracy. Our experiments carried out on several benchmarks show that our approach typically requires two orders of magnitude less memory than fastText while being only slightly inferior with respect to accuracy. As a result, it outperforms the state of the art by a good margin in terms of the compromise between memory usage and accuracy. ",
+        "bbox": [
+            233,
+            281,
+            764,
+            460
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 INTRODUCTION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            489,
+            336,
+            505
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text classification is an important problem in Natural Language Processing (NLP). Real world usecases include spam filtering or e-mail categorization. It is a core component in more complex systems such as search and ranking. Recently, deep learning techniques based on neural networks have achieved state of the art results in various NLP applications. One of the main successes of deep learning is due to the effectiveness of recurrent networks for language modeling and their application to speech recognition and machine translation (Mikolov, 2012). However, in other cases including several text classification problems, it has been shown that deep networks do not convincingly beat the prior state of the art techniques (Wang & Manning, 2012; Joulin et al., 2016). ",
+        "bbox": [
+            174,
+            521,
+            825,
+            632
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In spite of being (typically) orders of magnitude slower to train than traditional techniques based on n-grams, neural networks are often regarded as a promising alternative due to compact model sizes, in particular for character based models. This is important for applications that need to run on systems with limited memory such as smartphones. ",
+        "bbox": [
+            174,
+            640,
+            825,
+            695
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This paper specifically addresses the compromise between classification accuracy and the model size. We extend our previous work implemented in the fastText library1. It is based on n-gram features, dimensionality reduction, and a fast approximation of the softmax classifier (Joulin et al., 2016). We show that a few key ingredients, namely feature pruning, quantization, hashing, and retraining, allow us to produce text classification models with tiny size, often less than 100kB when trained on several popular datasets, without noticeably sacrificing accuracy or speed. ",
+        "bbox": [
+            174,
+            703,
+            823,
+            786
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We plan to publish the code and scripts required to reproduce our results as an extension of the fastText library, thereby providing strong reproducible baselines for text classifiers that optimize the compromise between the model size and accuracy. We hope that this will help the engineering community to improve existing applications by using more efficient models. ",
+        "bbox": [
+            176,
+            794,
+            825,
+            849
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "This paper is organized as follows. Section 2 introduces related work, Section 3 describes our text classification model and explains how we drastically reduce the model size. Section 4 shows the effectiveness of our approach in experiments on multiple text classification benchmarks. ",
+        "bbox": [
+            176,
+            856,
+            823,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 RELATED WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            339,
+            117
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Models for text classification. Text classification is a problem that has its roots in many applications such as web search, information retrieval and document classification (Deerwester et al., 1990; Pang & Lee, 2008). Linear classifiers often obtain state-of-the-art performance while being scalable (Agarwal et al., 2014; Joachims, 1998; Joulin et al., 2016; McCallum & Nigam, 1998). They are particularly interesting when associated with the right features (Wang & Manning, 2012). They usually require storing embeddings for words and n-grams, which makes them memory inefficient. ",
+        "bbox": [
+            174,
+            133,
+            825,
+            217
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Compression of language models. Our work is related to compression of statistical language models. Classical approaches include feature pruning based on entropy (Stolcke, 2000) and quantization. Pruning aims to keep only the most important n-grams in the model, leaving out those with probability lower than a specified threshold. Further, the individual n-grams can be compressed by quantizing the probability value, and by storing the n-gram itself more efficiently than as a sequence of characters. Various strategies have been developed, for example using tree structures or hash functions, and are discussed in (Talbot & Brants, 2008). ",
+        "bbox": [
+            174,
+            233,
+            825,
+            330
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Compression for similarity estimation and search. There is a large body of literature on how to compress a set of vectors into compact codes, such that the comparison of two codes approximates a target similarity in the original space. The typical use-case of these methods considers an indexed dataset of compressed vectors, and a query for which we want to find the nearest neighbors in the indexed set. One of the most popular is Locality-sensitive hashing (LSH) by Charikar (2002), which is a binarization technique based on random projections that approximates the cosine similarity between two vectors through a monotonous function of the Hamming distance between the two corresponding binary codes. In our paper, LSH refers to this binarization strategy2. Many subsequent works have improved this initial binarization technique, such as spectal hashing (Weiss et al., 2009), or Iterative Quantization (ITQ) (Gong & Lazebnik, 2011), which learns a rotation matrix minimizing the quantization loss of the binarization. We refer the reader to two recent surveys by Wang et al. (2014) and Wang et al. (2015) for an overview of the binary hashing literature. ",
+        "bbox": [
+            174,
+            347,
+            825,
+            513
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Beyond these binarization strategies, more general quantization techniques derived from Jegou et al. (2011) offer better trade-offs between memory and the approximation of a distance estimator. The Product Quantization (PQ) method approximates the distances by calculating, in the compressed domain, the distance between their quantized approximations. This method is statistically guaranteed to preserve the Euclidean distance between the vectors within an error bound directly related to the quantization error. The original PQ has been concurrently improved by Ge et al. (2013) and Norouzi & Fleet (2013), who learn an orthogonal transform minimizing the overall quantization loss. In our paper, we will consider the Optimized Product Quantization (OPQ) variant (Ge et al., 2013). ",
+        "bbox": [
+            174,
+            521,
+            825,
+            632
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Softmax approximation The aforementioned works approximate either the Euclidean distance or the cosine similarity (both being equivalent in the case of unit-norm vectors). However, in the context of fastText, we are specifically interested in approximating the maximum inner product involved in a softmax layer. Several approaches derived from LSH have been recently proposed to achieve this goal, such as Asymmetric LSH by Shrivastava & Li (2014), subsequently discussed by Neyshabur & Srebro (2015). In our work, since we are not constrained to purely binary codes, we resort a more traditional encoding by employing a magnitude/direction parametrization of our vectors. Therefore we only need to encode/compress an unitary d-dimensional vector, which fits the aforementioned LSH and PQ methods well. ",
+        "bbox": [
+            174,
+            647,
+            825,
+            773
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Neural network compression models. Recently, several research efforts have been conducted to compress the parameters of architectures involved in computer vision, namely for state-of-theart Convolutional Neural Networks (CNNs) (Han et al., 2016; Lin et al., 2015). Some use vector quantization (Gong et al., 2014) while others binarize the network (Courbariaux et al., 2016). Denil et al. (2013) show that such classification models are easily compressed because they are overparametrized, which concurs with early observations by LeCun et al. (1990). ",
+        "bbox": [
+            174,
+            790,
+            825,
+            873
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Some of these works both aim at reducing the model size and the speed. In our case, since the fastText classifier on which our proposal is built upon is already very efficient, we are primilarly interested in reducing the size of the model while keeping a comparable classification efficiency. ",
+        "bbox": [
+            174,
+            103,
+            825,
+            146
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 PROPOSED APPROACH",
+        "text_level": 1,
+        "bbox": [
+            176,
+            166,
+            390,
+            183
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 TEXT CLASSIFICATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            196,
+            372,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In the context of text classification, linear classifiers (Joulin et al., 2016) remain competitive with more sophisticated, deeper models, and are much faster to train. On top of standard tricks commonly used in linear text classification (Agarwal et al., 2014; Wang & Manning, 2012; Weinberger et al., 2009), Joulin et al. (2016) use a low rank constraint to reduce the computation burden while sharing information between different classes. This is especially useful in the case of a large output space, where rare classes may have only a few training examples. In this paper, we focus on a similar model, that is, which minimizes the softmax loss $\\ell$ over $N$ documents: ",
+        "bbox": [
+            173,
+            222,
+            825,
+            320
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/c05cd0c78ee183bfdc3d491b82c10824be5f6171e8f5cd0618e076f92ec5d0d3.jpg",
+        "text": "$$\n\\sum _ { n = 1 } ^ { N } \\ell ( y _ { n } , B A x _ { n } ) ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            419,
+            325,
+            542,
+            368
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where $x _ { n }$ is a bag of one-hot vectors and $y _ { n }$ the label of the $n$ -th document. In the case of a large vocabulary and a large output space, the matrices $A$ and $B$ are big and can require gigabytes of memory. Below, we describe how we reduce this memory usage. ",
+        "bbox": [
+            174,
+            372,
+            825,
+            415
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 BOTTOM-UP PRODUCT QUANTIZATION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            431,
+            480,
+            445
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Product quantization is a popular method for compressed-domain approximate nearest neighbor search (Jegou et al., 2011). As a compression technique, it approximates a real-valued vector by finding the closest vector in a pre-defined structured set of centroids, referred to as a codebook. This codebook is not enumerated, since it is extremely large. Instead it is implicitly defined by its structure: a $d$ -dimensional vector $x \\in \\mathbb { R } ^ { d }$ is approximated as ",
+        "bbox": [
+            174,
+            457,
+            825,
+            527
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/da90bf504a17fc52277df137768f58020e30b1c0175980506fa045aa6bb8dafe.jpg",
+        "text": "$$\n\\hat { x } = \\sum _ { i = 1 } ^ { k } q _ { i } ( x ) ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            446,
+            532,
+            547,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where the different subquantizers $q _ { i } : x \\mapsto q _ { i } ( x )$ are complementary in the sense that their respective centroids lie in distinct orthogonal subspaces, i.e., $\\forall i \\neq j$ , $\\forall x , y$ , $\\langle q _ { i } ( x ) | q _ { j } ( y ) \\rangle = 0$ . In the original PQ, the subspaces are aligned with the natural axis, while OPQ learns a rotation, which amounts to alleviating this constraint and to not depend on the original coordinate system. Another way to see this is to consider that PQ splits a given vector $x$ into $k$ subvectors $x ^ { i }$ , $i \\stackrel { \\cdot } { = } 1 \\ldots k$ , each of dimension $d / k$ : $x = [ x ^ { 1 } \\dots x ^ { i } \\dots x ^ { k } ]$ , and quantizes each sub-vector using a distinct $\\mathbf { k }$ -means quantizer. Each subvector $x ^ { i }$ is thus mapped to the closest centroid amongst $2 ^ { b }$ centroids, where $b$ is the number of bits required to store the quantization index of the subquantizer, typically $b = 8$ . The reconstructed vector can take $2 ^ { k b }$ distinct reproduction values, and is stored in $k b$ bits. ",
+        "bbox": [
+            173,
+            580,
+            825,
+            707
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "PQ estimates the inner product in the compressed domain as ",
+        "bbox": [
+            173,
+            713,
+            570,
+            728
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "img_path": "images/02fad22067c82eb0dad378962f56e449cf2de483837478a0d1407957fcdf283b.jpg",
+        "text": "$$\n\\boldsymbol { x } ^ { \\top } \\boldsymbol { y } \\approx \\hat { \\boldsymbol { x } } ^ { \\top } \\boldsymbol { y } = \\sum _ { i = 1 } ^ { k } q _ { i } ( \\boldsymbol { x } ^ { i } ) ^ { \\top } \\boldsymbol { y } ^ { i } .\n$$",
+        "text_format": "latex",
+        "bbox": [
+            398,
+            733,
+            599,
+            777
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This is a straightforward extension of the square L2 distance estimation of Jegou et al. (2011). In practice, the vector estimate $\\hat { x }$ is trivially reconstructed from the codes, i.e., from the quantization indexes, by concatenating these centroids. ",
+        "bbox": [
+            173,
+            789,
+            825,
+            832
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The two parameters involved in PQ, namely the number of subquantizers $k$ and the number of bits $b$ per quantization index, are typically set to $\\dot { k } \\in [ 2 , d / 2 ]$ , and $b = 8$ to ensure byte-alignment. ",
+        "bbox": [
+            173,
+            838,
+            823,
+            867
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Discussion. PQ offers several interesting properties in our context of text classification. Firstly, the training is very fast because the subquantizers have a small number of centroids, i.e., 256 centroids for $b = 8$ . Secondly, at test time it allows the reconstruction of the vectors with almost no computational and memory overhead. Thirdly, it has been successfully applied in computer vision, offering much better performance than binary codes, which makes it a natural candidate to compress relatively shallow models. As observed by Sanchez & Perronnin (2011), using PQ just before the ´ last layer incurs a very limited loss in accuracy when combined with a support vector machine. ",
+        "bbox": [
+            176,
+            881,
+            823,
+            924
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "",
+        "bbox": [
+            174,
+            103,
+            823,
+            159
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the context of text classification, the norms of the vectors are widely spread, typically with a ratio of 1000 between the max and the min. Therefore kmeans performs poorly because it optimizes an absolute error objective, so it maps all low-norm vectors to 0. A simple solution is to separate the norm and the angle of the vectors and to quantize them separately. This allows a quantization with no loss of performance, yet requires an extra $b$ bits per vector. ",
+        "bbox": [
+            174,
+            166,
+            825,
+            236
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Bottom-up strategy: re-training. The first works aiming at compressing CNN models like the one proposed by (Gong et al., 2014) used the reconstruction from off-the-shelf PQ, i.e., without any re-training. However, as observed in Sablayrolles et al. (2016), when using quantization methods like PQ, it is better to re-train the layers occurring after the quantization, so that the network can re-adjust itself to the quantization. There is a strong argument arguing for this re-training strategy: the square magnitude of vectors is reduced, on average, by the average quantization error for any quantizer satisfying the Lloyd conditions; see Jegou et al. (2011) for details. ",
+        "bbox": [
+            174,
+            251,
+            825,
+            349
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "This suggests a bottom-up learning strategy where we first quantize the input matrix, then retrain and quantize the output matrix (the input matrix being frozen). Experiments in section 4 show that it is worth adopting this strategy. ",
+        "bbox": [
+            174,
+            357,
+            825,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Memory savings with PQ. In practice, the bottom-up PQ strategy offers a compression factor of 10 without any noticeable loss of performance. Without re-training, we notice a drop in accuracy between $0 . 1 \\%$ and $0 . 5 \\%$ , depending on the dataset and setting; see Section 4 and the appendix. ",
+        "bbox": [
+            176,
+            414,
+            825,
+            457
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 FURTHER TEXT SPECIFIC TRICKS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            473,
+            442,
+            488
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The memory usage strongly depends on the size of the vocabulary, which can be large in many text classification tasks. While it is clear that a large part of the vocabulary is useless or redundant, directly reducing the vocabulary to the most frequent words is not satisfactory: most of the frequent words, like “the” or “is” are not discriminative, in contrast to some rare words, e.g., in the context of tag prediction. In this section, we discuss a few heuristics to reduce the space taken by the dictionary. They lead to major memory reduction, in extreme cases by a factor 100. We experimentally show that this drastic reduction is complementary with the PQ compression method, meaning that the combination of both strategies reduces the model size by a factor up to $\\times 1 0 0 0$ for some datasets. ",
+        "bbox": [
+            174,
+            500,
+            825,
+            611
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pruning the vocabulary. Discovering which word or n-gram must be kept to preserve the overall performance is a feature selection problem. While many approaches have been proposed to select groups of variables during training (Bach et al., 2012; Meier et al., 2008), we are interested in selecting a fixed subset of $K$ words and ngrams from a pre-trained model. This can be achieved by selecting the $K$ embeddings that preserve as much of the model as possible, which can be reduced to selecting the $K$ words and ngrams associated with the highest norms. ",
+        "bbox": [
+            174,
+            626,
+            825,
+            710
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "While this approach offers major memory savings, it has one drawback occurring in some particular cases: some documents may not contained any of the $K$ best features, leading to a significant drop in performance. It is thus important to keep the $K$ best features under the condition that they cover the whole training set. More formally, the problem is to find a subset $s$ in the feature set $\\nu$ that maximizes the sum of their norms $w _ { s }$ under the constraint that all the documents in the training set $\\mathcal { D }$ are covered: ",
+        "bbox": [
+            174,
+            717,
+            825,
+            800
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "img_path": "images/130acb8ea66ae03ff25174aa0c7531c2bda761c18bb70a3b1cef1cb6635a9218.jpg",
+        "text": "$$\n\\operatorname* { m a x } _ { S \\subseteq \\mathcal { V } } \\sum _ { s \\in \\cal { S } } w _ { s } \\quad \\mathrm { ~ s . t . ~ } \\quad | S | \\leq K , P 1 _ { \\cal { S } } \\geq 1 _ { \\cal { D } } ,\n$$",
+        "text_format": "latex",
+        "bbox": [
+            341,
+            804,
+            651,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where $P$ is a matrix such that $P _ { d s } = 1$ if the $s$ -th feature is in the $d$ -th document, and 0 otherwise. This problem is directly related to set covering problems that are NP-hard (Feige, 1998). Standard greedy approaches require the storing of an inverted index or to do multiple passes over the dataset, which is prohibitive on very large dataset (Chierichetti et al., 2010). This problem can be cast as an instance of online submodular maximization with a rank constraint (Badanidiyuru et al., 2014; ",
+        "bbox": [
+            174,
+            853,
+            825,
+            924
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/15d78f88290b96691cc9018373eef6a0532bbee755b0815e5eb2080d3353a831.jpg",
+        "image_caption": [
+            "Figure 1: Accuracy as a function of the memory per vector/embedding on 3 datasets from Zhang et al. (2015). Note, an extra byte is required when we encode the norm explicitly (”norm”). "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            184,
+            107,
+            808,
+            295
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Bateni et al., 2010). In our case, we use a simple online parallelizable greedy approach: For each document, we verify if it is already covered by a retained feature and, if not, we add the feature with the highest norm to our set of retained features. If the number of features is below $k$ , we add the features with the highest norm that have not yet been picked. ",
+        "bbox": [
+            174,
+            387,
+            825,
+            444
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hashing trick & Bloom filter. On small models, the dictionary can take a significant portion of the memory. Instead of saving it, we extend the hashing trick used in Joulin et al. (2016) to both words and n-grams. This strategy is also used in Vowpal Wabbit (Agarwal et al., 2014) in the context of online training. This allows us to save around 1-2Mb with almost no overhead at test time (just the cost of computing the hashing function). ",
+        "bbox": [
+            174,
+            478,
+            825,
+            549
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Pruning the vocabulary while using the hashing trick requires keeping a list of the indices of the $K$ remaining buckets. At test time, a binary search over the list of indices is required. It has a complexity of $O ( \\log ( K ) )$ and a memory overhead of a few hundreds of kilobytes. Using Bloom filters instead reduces the complexity $\\mathcal { O } ( 1 )$ at test time and saves a few hundred kilobytes. However, in practice, it degrades performance. ",
+        "bbox": [
+            174,
+            555,
+            825,
+            626
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 EXPERIMENTS ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            665,
+            326,
+            681
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This section evaluates the quality of our model compression pipeline and compare it to other compression methods on different text classification problems, and to other compact text classifiers. ",
+        "bbox": [
+            174,
+            707,
+            821,
+            736
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation protocol and datasets. Our experimental pipeline is as follows: we train a model using fastText with the default setting unless specified otherwise. That is 2M buckets, a learning rate of 0.1 and 10 training epochs. The dimensionality $d$ of the embeddings is set to powers of 2 to avoid border effects that could make the interpretation of the results more difficult. As baselines, we use Locality-Sensitive Hashing (LSH) (Charikar, 2002), PQ (Jegou et al., 2011) and OPQ (Ge et al., 2013) (the non-parametric variant). Note that we use an improved version of LSH where random orthogonal matrices are used instead of random matrix projection Jegou et al. (2008). In a first ´ series of experiments, we use the 8 datasets and evaluation protocol of Zhang et al. (2015). These datasets contain few million documents and have at most 10 classes. We also explore the limit of quantization on a dataset with an extremely large output space, that is a tag dataset extracted from the YFCC100M collection (Thomee et al., 2016)3, referred to as FlickrTag in the rest of this paper. ",
+        "bbox": [
+            174,
+            770,
+            825,
+            924
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/cdcd3e63a3059e96131d7fa08c0a03f66afb2540d2d14097ce9d24bf7f2608de.jpg",
+        "image_caption": [
+            "Figure 2: Loss of accuracy as a function of the model size. We compare the compressed model with different level of pruning with NPQ and the full fastText model. We also compare with Zhang et al. (2015) and Xiao & Cho (2016). Note that the size is in log scale. ",
+            "Table 1: Best ranked words w.r.t. entropy (left) and norm (right) on the Amazon full review dataset. We give the rank for both criteria. The norm ranking filters out words carrying little information. "
+        ],
+        "image_footnote": [],
+        "bbox": [
+            215,
+            107,
+            779,
+            429
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 SMALL DATASETS ",
+        "text_level": 1,
+        "bbox": [
+            174,
+            507,
+            338,
+            521
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Compression techniques. We compare three popular methods used for similarity estimation with compact codes: LSH, PQ and OPQ on the datasets released by Zhang et al. (2015). Figure 1 shows the accuracy as a function of the number of bytes used per embedding, which corresponds to the number $k$ of subvectors in the case of PQ and OPQ. See more results in the appendix. As discussed in Section 2, LSH reproduces the cosine similarity and is therefore not adapted to un-normalized data. Therefore we only report results with normalization. Once normalized, PQ and OPQ are almost lossless even when using only $k = 4$ subquantizers per embedding (equivalently, bytes). We observe in practice that using $k = d / 2$ , i.e., half of the components of the embeddings, works well in practice. In the rest of the paper and if not stated otherwise, we focus on this setting. The difference between the normalized versions of PQ and OPQ is limited and depends on the dataset. Therefore we adopt the normalized PQ (NPQ) for the rest of this study, since it is faster to train. ",
+        "bbox": [
+            173,
+            532,
+            825,
+            685
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/5a6d1d3a78e17847d7f942fce2c5a8c91548f3864931f56d56e31adf1e251089.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>word</td><td>Entropy</td><td>Norm</td><td>word</td><td>Entropy</td><td>Norm</td></tr><tr><td>·</td><td>1</td><td>354</td><td>mediocre</td><td>1399</td><td>1</td></tr><tr><td>，</td><td>2</td><td>176</td><td>disappointing</td><td>454</td><td>2</td></tr><tr><td>the</td><td>3</td><td>179</td><td>SO-SO</td><td>2809</td><td>3</td></tr><tr><td>and</td><td>4</td><td>1639</td><td>lacks</td><td>1244</td><td>4</td></tr><tr><td>i</td><td>5</td><td>2374</td><td>worthless</td><td>1757</td><td>5</td></tr><tr><td>a</td><td>6</td><td>970</td><td>dreadful</td><td>4358</td><td>6</td></tr><tr><td>to</td><td>7</td><td>1775</td><td>drm</td><td>6395</td><td>7</td></tr><tr><td>it</td><td>8</td><td>1956</td><td>poorly</td><td>716</td><td>8</td></tr><tr><td>of</td><td>9</td><td>2815</td><td>uninspired</td><td>4245</td><td>9</td></tr><tr><td>this</td><td>10</td><td>3275</td><td>worst</td><td>402</td><td>10</td></tr></table>",
+        "bbox": [
+            289,
+            699,
+            709,
+            853
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/b88d49372c0f39d2c472eb7c2d26f438cecfa02b2501a0ef991d696cfaa1c7c3.jpg",
+        "table_caption": [
+            "Table 2: Performance on very small models. We use a quantization with $k = 1$ , hashing and an extreme pruning. The last row shows the average drop of performance for different size. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>full</td><td>64KiB</td><td>32KiB</td><td>16 KiB</td></tr><tr><td>AG</td><td>65M</td><td>92.1 91.4</td><td>90.6</td><td>89.1</td></tr><tr><td>Amazon full</td><td>108M</td><td>60.0 58.8</td><td>56.0</td><td>52.9</td></tr><tr><td>Amazon pol.</td><td>113M</td><td>94.5 93.3</td><td>92.1</td><td>89.3</td></tr><tr><td>DBPedia</td><td>87M</td><td>98.4 98.2</td><td>98.1</td><td>97.4</td></tr><tr><td>Sogou</td><td>73M</td><td>96.4 96.4</td><td>96.3</td><td>95.5</td></tr><tr><td>Yahoo</td><td>122M</td><td>72.1 70.0</td><td>69.0</td><td>69.2</td></tr><tr><td>Yelp full</td><td>78M</td><td>63.8 63.2</td><td>62.4</td><td>58.7</td></tr><tr><td>Yelp pol.</td><td>77M</td><td>95.7 95.3</td><td>94.9</td><td>93.2</td></tr><tr><td>Average diff. [%]</td><td></td><td>0</td><td>-0.8 -1.7</td><td>-3.5</td></tr></table>",
+        "bbox": [
+            279,
+            101,
+            717,
+            262
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Pruning. Figure 2 shows the performance of our model with different sizes. We fix $k = d / 2$ and use different pruning thresholds. NPQ offers a compression rate of $\\times 1 0$ compared to the full model. As the pruning becomes more agressive, the overall compression can increase up up to $\\times 1 , 0 0 0$ with little drop of performance and no additional overhead at test time. In fact, using a smaller dictionary makes the model faster at test time. We also compare with character-level Convolutional Neural Networks (CNN) (Zhang et al., 2015; Xiao & Cho, 2016). They are attractive models for text classification because they achieve similar performance with less memory usage than linear models (Xiao & Cho, 2016). Even though fastText with the default setting uses more memory, NPQ is already on par with CNNs’ memory usage. Note that CNNs are not quantized, and it would be worth seeing how much they can be quantized with no drop of performance. Such a study is beyond the scope of this paper. Our pruning is based on the norm of the embeddings according to the guidelines of Section 3.3. Table 1 compares the ranking obtained with norms to the ranking obtained using entropy, which is commonly used in unsupervised settings Stolcke (2000). ",
+        "bbox": [
+            173,
+            325,
+            825,
+            507
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Extreme compression. Finally, in Table 2, we explore the limit of quantized model by looking at the performance obtained for models under 64KiB. Surprisingly, even at 64KiB and 32KiB, the drop of performance is only around $0 . 8 \\%$ and $1 . 7 \\%$ despite a compression rate of $\\times 1 , 0 0 0 - 4 , 0 0 0$ . ",
+        "bbox": [
+            174,
+            522,
+            825,
+            565
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.2 LARGE DATASET: FLICKRTAG ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            582,
+            421,
+            597
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In this section, we explore the limit of compression algorithms on very large datasets. Similar to Joulin et al. (2016), we consider a hashtag prediction dataset containing 312, 116 labels. We set the minimum count for words at 10, leading to a dictionary of 1, 427, 667 words. We take 10M buckets for n-grams and a hierarchical softmax. We refer to this dataset as FlickrTag. ",
+        "bbox": [
+            174,
+            608,
+            825,
+            664
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/9dd039811ceba7becce7e87300b0c7dfc5f9eecce1bcd7813cfa807b1261647a.jpg",
+        "table_caption": [
+            "Output encoding. We are interested in understanding how the performance degrades if the classifier is also quantized (i.e., the matrix $B$ in Eq. 1) and when the pruning is at the limit of the minimum number of features required to cover the full dataset. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>k</td><td>norm</td><td>retrain</td><td>Acc.</td><td>Size</td></tr><tr><td>full (uncompressed)</td><td></td><td></td><td></td><td>45.4</td><td>12 GiB</td></tr><tr><td>Input</td><td>128</td><td></td><td></td><td>45.0</td><td>1.7 GiB</td></tr><tr><td>Input</td><td>128</td><td>X</td><td></td><td>45.3</td><td>1.8 GiB</td></tr><tr><td>Input</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>1.8 GiB</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td></td><td>45.2</td><td>1.5 GiB</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td>X</td><td>45.4</td><td>1.5 GiB</td></tr></table>",
+        "bbox": [
+            281,
+            734,
+            715,
+            854
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: FlickrTag: Influence of quantizing the output matrix on performance. We use PQ for quantization with an optional normalization. We also retrain the output matrix after quantizing the input one. The ”norm” refers to the separate encoding of the magnitude and angle, while ”retrain” refers to the re-training bottom-up PQ method described in Section 3.2. ",
+        "bbox": [
+            173,
+            864,
+            825,
+            921
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows that quantizing both the “input” matrix (i.e., $A$ in Eq. 1) and the “output” matrix (i.e., $B$ ) does not degrade the performance compared to the full model. We use embeddings with $d = 2 5 6$ dimensions and use $k = d / 2$ subquantizers. We do not use any text specific tricks, which leads to a compression factor of 8. Note that even if the output matrix is not retrained over the embeddings, the performance is only $0 . 2 \\%$ away from the full model. As shown in the Appendix, using less subquantizers significantly decreases the performance for a small memory gain. ",
+        "bbox": [
+            174,
+            103,
+            825,
+            188
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/0f5d3e973dd3a009925f06836fbc45a469f92c4d8ba04b6106d5c3063189683b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>full</td><td colspan=\"2\"> Entropy pruning</td><td colspan=\"2\">Norm pruning</td><td colspan=\"2\">Max-Cover pruning</td></tr><tr><td>#embeddings Memory</td><td>11.5M 12GiB</td><td>2M 297MiB</td><td>1M 174MiB</td><td>2M 305MiB</td><td>1M 179MiB</td><td>2M 305MiB</td><td>1M 179MiB</td></tr><tr><td>Coverage [%]</td><td>88.4</td><td>70.5</td><td>70.5</td><td>73.2</td><td>61.9</td><td>88.4</td><td>88.4</td></tr><tr><td>Accuracy</td><td>45.4</td><td>32.1</td><td>30.5</td><td>41.6</td><td>35.8</td><td>45.5</td><td>43.9</td></tr></table>",
+        "bbox": [
+            186,
+            202,
+            810,
+            295
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: FlickrTag: Comparison of entropy pruning, norm pruning and max-cover pruning methods.   \nWe show the coverage of the test set for each method. ",
+        "bbox": [
+            174,
+            304,
+            821,
+            333
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Pruning. Table 4 shows how the performance evolves with pruning. We measure this effect on top of a fully quantized model. The full model misses $1 1 . 6 \\%$ of the test set because of missing words (some documents are either only composed of hashtags or have only rare words). There are 312, 116 labels and thus it seems reasonable to keep embeddings in the order of the million. A naive pruning with 1M features misses about $3 0 - 4 0 \\%$ of the test set, leading to a significant drop of performance. On the other hand, even though the max-coverage pruning approach was set on the train set, it does not suffer from any coverage loss on the test set. This leads to a smaller drop of performance. If the pruning is too aggressive, however, the coverage decreases significantly. ",
+        "bbox": [
+            173,
+            359,
+            825,
+            472
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 FUTURE WORK ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            492,
+            334,
+            508
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It may be possible to obtain further reduction of the model size in the future. One idea is to condition the size of the vectors (both for the input features and the labels) based on their frequency (Chen et al., 2015; Grave et al., 2016). For example, it is probably not worth representing the rare labels by full 256-dimensional vectors in the case of the FlickrTag dataset. Thus, conditioning the vector size on the frequency and norm seems like an interesting direction to explore in the future. ",
+        "bbox": [
+            174,
+            526,
+            825,
+            597
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We may also consider combining the entropy and norm pruning criteria: instead of keeping the features in the model based just on the frequency or the norm, we can use both to keep a good set of features. This could help to keep features that are both frequent and discriminative, and thereby to reduce the coverage problem that we have observed. ",
+        "bbox": [
+            174,
+            603,
+            825,
+            659
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Additionally, instead of pruning out the less useful features, we can decompose them into smaller units (Mikolov et al., 2012). For example, this can be achieved by splitting every non-discriminative word into a sequence of character trigrams. This could help in cases where training and test examples are very short (for example just a single word). ",
+        "bbox": [
+            174,
+            665,
+            825,
+            722
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 CONCLUSION ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            744,
+            318,
+            761
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we have presented several simple techniques to reduce, by several orders of magnitude, the memory complexity of certain text classifiers without sacrificing accuracy nor speed. This is achieved by applying discriminative pruning which aims to keep only important features in the trained model, and by performing quantization of the weight matrices and hashing of the dictionary. ",
+        "bbox": [
+            174,
+            776,
+            825,
+            833
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We will publish the code as an extension of the fastText library. We hope that our work will serve as a baseline to the research community, where there is an increasing interest for comparing the performance of various deep learning text classifiers for a given number of parameters. Overall, compared to recent work based on convolutional neural networks, fastText.zip is often more accurate, while requiring several orders of magnitude less time to train on common CPUs, and incurring a fraction of the memory complexity. ",
+        "bbox": [
+            174,
+            840,
+            825,
+            924
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "REFERENCES ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            102,
+            287,
+            118
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Alekh Agarwal, Olivier Chapelle, Miroslav Dud´ık, and John Langford. A reliable effective terascale linear learning system. Journal of Machine Learning Research, 15(1):1111–1133, 2014. ",
+        "bbox": [
+            174,
+            126,
+            823,
+            155
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Francis Bach, Rodolphe Jenatton, Julien Mairal, and Guillaume Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends $\\textsuperscript { \\textregistered }$ in Machine Learning, 4(1):1–106, 2012. ",
+        "bbox": [
+            173,
+            162,
+            823,
+            193
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ashwinkumar Badanidiyuru, Baharan Mirzasoleiman, Amin Karbasi, and Andreas Krause. Streaming submodular maximization: Massive data summarization on the fly. In SIGKDD, pp. 671–680. ACM, 2014. ",
+        "bbox": [
+            173,
+            200,
+            823,
+            242
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Mohammad Hossein Bateni, Mohammad Taghi Hajiaghayi, and Morteza Zadimoghaddam. Submodular secretary problem and extensions. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pp. 39–52. Springer, 2010. ",
+        "bbox": [
+            173,
+            252,
+            823,
+            295
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Moses S. Charikar. Similarity estimation techniques from rounding algorithms. In STOC, pp. 380– 388, May 2002. ",
+        "bbox": [
+            171,
+            304,
+            823,
+            333
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Welin Chen, David Grangier, and Michael Auli. Strategies for training large vocabulary neural language models. arXiv preprint arXiv:1512.04906, 2015. ",
+        "bbox": [
+            169,
+            340,
+            823,
+            371
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Flavio Chierichetti, Ravi Kumar, and Andrew Tomkins. Max-cover in map-reduce. In International Conference on World Wide Web, 2010. ",
+        "bbox": [
+            171,
+            378,
+            823,
+            409
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized neural networks: Training neural networks with weights and activations constrained to $+ 1$ or $^ { - 1 }$ . arXiv preprint arXiv:1602.02830, 2016. ",
+        "bbox": [
+            174,
+            416,
+            825,
+            459
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "M. Datar, N. Immorlica, P. Indyk, and V.S. Mirrokni. Locality-sensitive hashing scheme based on pstable distributions. In Proceedings of the Symposium on Computational Geometry, pp. 253–262, 2004. ",
+        "bbox": [
+            173,
+            467,
+            823,
+            510
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. Indexing by latent semantic analysis. Journal of the American society for information science, 1990. ",
+        "bbox": [
+            173,
+            518,
+            825,
+            561
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Misha Denil, Babak Shakibi, Laurent Dinh, Marc-Aurelio Ranzato, and Nando et all de Freitas. Predicting parameters in deep learning. In NIPS, pp. 2148–2156, 2013. ",
+        "bbox": [
+            176,
+            570,
+            823,
+            601
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Uriel Feige. A threshold of ln n for approximating set cover. JACM, 45(4):634–652, 1998. ",
+        "bbox": [
+            174,
+            608,
+            766,
+            625
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Tiezheng Ge, Kaiming He, Qifa Ke, and Jian Sun. Optimized product quantization for approximate nearest neighbor search. In CVPR, June 2013. ",
+        "bbox": [
+            174,
+            631,
+            820,
+            661
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yunchao Gong and Svetlana Lazebnik. Iterative quantization: A procrustean approach to learning binary codes. In CVPR, June 2011. ",
+        "bbox": [
+            174,
+            670,
+            821,
+            699
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Yunchao Gong, Liu Liu, Ming Yang, and Lubomir Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. ",
+        "bbox": [
+            174,
+            707,
+            823,
+            736
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Edouard Grave, Armand Joulin, Moustapha Cisse, David Grangier, and Herv ´ e J ´ egou. Efficient ´ softmax approximation for gpus. arXiv preprint arXiv:1609.04309, 2016. ",
+        "bbox": [
+            174,
+            744,
+            825,
+            775
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. ",
+        "bbox": [
+            174,
+            781,
+            825,
+            811
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Herve J ´ egou, Matthijs Douze, and Cordelia Schmid. Hamming embedding and weak geometric ´ consistency for large scale image search. In ECCV, October 2008. ",
+        "bbox": [
+            173,
+            820,
+            821,
+            849
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor ´ search. IEEE Trans. PAMI, January 2011. ",
+        "bbox": [
+            174,
+            857,
+            821,
+            887
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Thorsten Joachims. Text categorization with support vector machines: Learning with many relevant features. Springer, 1998. ",
+        "bbox": [
+            174,
+            895,
+            821,
+            924
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. ",
+        "bbox": [
+            171,
+            103,
+            825,
+            132
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yann LeCun, John S Denker, and Sara A Solla. Optimal brain damage. NIPS, 2:598–605, 1990. ",
+        "bbox": [
+            176,
+            140,
+            805,
+            156
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Zhouhan Lin, Matthieu Courbariaux, Roland Memisevic, and Yoshua Bengio. Neural networks with few multiplications. arXiv preprint arXiv:1510.03009, 2015. ",
+        "bbox": [
+            176,
+            164,
+            823,
+            194
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Andrew McCallum and Kamal Nigam. A comparison of event models for naive bayes text classification. In AAAI workshop on learning for text categorization, 1998. ",
+        "bbox": [
+            176,
+            200,
+            821,
+            231
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Lukas Meier, Sara Van De Geer, and Peter Buhlmann. The group lasso for logistic regression. ¨ Journal of the Royal Statistical Society: Series B (Statistical Methodology), 70(1):53–71, 2008. ",
+        "bbox": [
+            174,
+            238,
+            821,
+            268
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Tomas Mikolov. Statistical language models based on neural networks. In PhD thesis. VUT Brno, 2012. ",
+        "bbox": [
+            174,
+            275,
+            823,
+            304
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Tomas Mikolov, Ilya Sutskever, Anoop Deoras, Hai-Son Le, Stefan Kombrink, and J Cernocky. Subword language modeling with neural networks. preprint, 2012. ",
+        "bbox": [
+            173,
+            311,
+            823,
+            342
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Behnam Neyshabur and Nathan Srebro. On symmetric and asymmetric lshs for inner product search. In ICML, pp. 1926–1934, 2015. ",
+        "bbox": [
+            173,
+            349,
+            823,
+            378
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Mohammad Norouzi and David Fleet. Cartesian k-means. In CVPR, June 2013. ",
+        "bbox": [
+            173,
+            386,
+            699,
+            402
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations and trends in information retrieval, 2008. ",
+        "bbox": [
+            174,
+            410,
+            823,
+            439
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Alexandre Sablayrolles, Matthijs Douze, Herve J ´ egou, and Nicolas Usunier. How should we evalu-´ ate supervised hashing? arXiv preprint arXiv:1609.06753, 2016. ",
+        "bbox": [
+            173,
+            446,
+            821,
+            477
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale im- ´ age classification. In CVPR, 2011. ",
+        "bbox": [
+            173,
+            484,
+            823,
+            513
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Anshumali Shrivastava and Ping Li. Asymmetric LSH for sublinear time maximum inner product search. In NIPS, pp. 2321–2329, 2014. ",
+        "bbox": [
+            173,
+            521,
+            821,
+            551
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Andreas Stolcke. Entropy-based pruning of backoff language models. arXiv preprint cs/0006025, 2000. ",
+        "bbox": [
+            173,
+            558,
+            823,
+            588
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "David Talbot and Thorsten Brants. Randomized language models via perfect hash functions. In ACL, 2008. ",
+        "bbox": [
+            173,
+            595,
+            823,
+            626
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In Communications of the ACM, 2016. ",
+        "bbox": [
+            173,
+            632,
+            823,
+            676
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jingdong Wang, Heng Tao Shen, Jingkuan Song, and Jianqiu Ji. Hashing for similarity search: A survey. arXiv preprint arXiv:1408.2927, 2014. ",
+        "bbox": [
+            171,
+            684,
+            821,
+            713
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Jun Wang, Wei Liu, Sanjiv Kumar, and Shih-Fu Chang. Learning to hash for indexing big data - A survey. CoRR, abs/1509.05472, 2015. ",
+        "bbox": [
+            171,
+            720,
+            823,
+            751
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Sida Wang and Christopher D Manning. Baselines and bigrams: Simple, good sentiment and topic classification. In ACL, 2012. ",
+        "bbox": [
+            173,
+            757,
+            823,
+            787
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Kilian Q Weinberger, Anirban Dasgupta, John Langford, Alex Smola, and Josh Attenberg. Feature hashing for large scale multitask learning. In ICML, 2009. ",
+        "bbox": [
+            173,
+            795,
+            823,
+            825
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yair Weiss, Antonio Torralba, and Rob Fergus. Spectral hashing. In NIPS, December 2009. ",
+        "bbox": [
+            171,
+            832,
+            776,
+            848
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Yijun Xiao and Kyunghyun Cho. Efficient character-level document classification by combining convolution and recurrent layers. arXiv preprint arXiv:1602.00367, 2016. ",
+        "bbox": [
+            171,
+            856,
+            821,
+            886
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In NIPS, 2015. ",
+        "bbox": [
+            171,
+            892,
+            823,
+            922
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "APPENDIX ",
+        "text_level": 1,
+        "bbox": [
+            176,
+            103,
+            263,
+            117
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/a2a6acc82590e8292c5be9f305174fbbbb2706328e9ca6446c5fa0412a92a47d.jpg",
+        "table_caption": [
+            "In the appendix, we show some additional results. The model used in these experiments only had 1M ngram buckets. In Table 5, we show a thorough comparison of LSH, PQ and OPQ on 8 different datasets. Table 7 summarizes the comparison with CNNs in terms of accuracy and size. Table 8 show a thorough comparison of the hashing trick and the Bloom filters. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Quant. k</td><td>ULIOU</td><td></td><td>AG</td><td></td><td>Amz. f.</td><td></td><td>Amz. p.</td><td></td><td>DBP</td><td></td><td></td><td>Sogou</td><td></td><td>Yah.</td><td></td><td>Yelp f.</td><td></td><td>Yelp p.</td></tr><tr><td>full full,nodict</td><td></td><td></td><td>92.1 36M 92.1 34M</td><td></td><td>59.897M 59.9</td><td>78M</td><td>94.5 94.5</td><td>104M 83M</td><td>98.467M 98.4</td><td>56M</td><td>96.347M 96.3</td><td>42M</td><td>72 72.2</td><td>120M 91M</td><td>63.756M 63.648M</td><td></td><td>95.7 53M 95.646M</td><td></td></tr><tr><td>LSH</td><td>8</td><td></td><td>88.7</td><td>8.5M</td><td>51.3</td><td>20M</td><td>90.3</td><td>21M</td><td>92.7</td><td>14M</td><td>94.2</td><td>11M</td><td>54.8</td><td>23M</td><td>56.7</td><td>12M</td><td>92.2</td><td>12M</td></tr><tr><td>PQ</td><td>8</td><td></td><td>91.7</td><td>8.5M</td><td>59.3</td><td>20M</td><td>94.4</td><td>21M</td><td>97.4</td><td>14M</td><td>96.1</td><td>11M</td><td>71.3</td><td>23M</td><td>62.8</td><td>12M</td><td>95.4</td><td>12M</td></tr><tr><td>OPQ</td><td>8</td><td></td><td>91.9</td><td>8.5M</td><td>59.3</td><td>20M</td><td>94.4</td><td>21M</td><td>96.9</td><td>14M</td><td>95.8</td><td>11M</td><td>71.4</td><td>23M</td><td>62.5</td><td>12M</td><td>95.4</td><td>12M</td></tr><tr><td>LSH</td><td>8</td><td>X</td><td>91.9</td><td>9.5M</td><td>59.4</td><td>22M</td><td>94.5</td><td>24M</td><td>97.8</td><td>16M</td><td>96.2</td><td>12M</td><td>71.6</td><td>26M</td><td>63.4</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>PQ</td><td>8</td><td>X</td><td>92.0</td><td>9.5M</td><td>59.8</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.1</td><td>26M</td><td>63.7</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>OPQ</td><td>8</td><td>X</td><td>92.1</td><td>9.5M</td><td>59.9</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.2</td><td>26M</td><td>63.6</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>LSH</td><td>4</td><td></td><td>88.3</td><td>4.3M</td><td>50.5</td><td>9.7M</td><td>88.9</td><td>11M</td><td>91.6</td><td>7.0M</td><td>94.3</td><td>5.3M</td><td>54.6</td><td>12M</td><td>56.5</td><td>6.0M</td><td>92.9</td><td>5.7M</td></tr><tr><td>PQ</td><td>4</td><td></td><td>91.6</td><td>4.3M</td><td>59.2</td><td>9.7M</td><td>94.4</td><td>11M</td><td>96.3</td><td>7.0M</td><td>96.1</td><td>5.3M</td><td>71.0</td><td>12M</td><td>62.2</td><td>6.0M</td><td>95.4</td><td>5.7M</td></tr><tr><td>OPQ</td><td>4</td><td></td><td>91.7</td><td>4.3M</td><td>59.0</td><td>9.7M</td><td>94.4</td><td>11M</td><td>96.9</td><td>7.0M</td><td>95.6</td><td>5.3M</td><td>71.2</td><td>12M</td><td>62.6</td><td>6.0M</td><td>95.4</td><td>5.7M</td></tr><tr><td>LSH</td><td>4</td><td>X</td><td>92.1</td><td>5.3M</td><td>59.2</td><td>13M</td><td>94.4</td><td>13M</td><td>97.7</td><td>8.8M</td><td>96.2</td><td>6.6M</td><td>71.1</td><td>15M</td><td>63.1</td><td>7.4M</td><td>95.5</td><td>7.2M</td></tr><tr><td>PQ</td><td>4</td><td>X</td><td>92.1</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.4</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72.0</td><td>15M</td><td>63.6</td><td>7.5M</td><td></td><td>95.67.2M</td></tr><tr><td>OPQ</td><td>4</td><td>X</td><td>92.2</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.3</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72.1</td><td>15M</td><td>63.7</td><td>7.5M</td><td>95.6</td><td>7.2M</td></tr><tr><td>LSH</td><td>2</td><td></td><td>87.7</td><td>2.2M</td><td>50.1</td><td>4.9M</td><td>88.9</td><td>5.2M</td><td>90.6</td><td>3.5M</td><td>93.9</td><td>2.7M</td><td>51.4</td><td>5.7M</td><td>56.6</td><td>3.0M</td><td>91.3</td><td>2.9M</td></tr><tr><td>PQ</td><td>2</td><td></td><td>91.1</td><td>2.2M</td><td>58.7</td><td>4.9M</td><td>94.4</td><td>5.2M</td><td>87.1</td><td>3.6M</td><td>95.3</td><td>2.7M</td><td>69.5</td><td>5.7M</td><td>62.1</td><td>3.0M</td><td>95.4</td><td>2.9M</td></tr><tr><td>OPQ</td><td>2</td><td></td><td>91.4</td><td>2.2M</td><td>58.2</td><td>4.9M</td><td>94.3</td><td>5.2M</td><td>91.6</td><td>3.6M</td><td>94.2</td><td>2.7M</td><td>69.6</td><td>5.7M</td><td>62.1</td><td>3.0M</td><td>95.4</td><td>2.9M</td></tr><tr><td>LSH</td><td>2</td><td>X</td><td>91.8</td><td>3.2M</td><td>58.6</td><td>7.3M</td><td>94.3</td><td>7.8M</td><td>97.1</td><td>5.3M</td><td>96.1</td><td>4.0M</td><td>69.7</td><td>8.6M</td><td>62.7</td><td>4.5M</td><td></td><td>95.54.3M</td></tr><tr><td>PQ</td><td>2</td><td>X</td><td>91.9</td><td>3.2M</td><td>59.6</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.3</td><td>4.0M</td><td>71.3</td><td>8.6M</td><td>63.44.5M</td><td></td><td></td><td>95.64.3M</td></tr><tr><td>OPQ</td><td>2</td><td>X</td><td>92.1</td><td>3.2M</td><td>59.5</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.2</td><td>4.0M</td><td>71.5</td><td>8.6M</td><td>63.44.5M</td><td></td><td></td><td>95.64.3M</td></tr></table>",
+        "bbox": [
+            75,
+            202,
+            910,
+            517
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/8bc40cec7313fcf1eca2cbf124b392f0266ca4e258655ad146c83644e83ca7ec.jpg",
+        "table_caption": [
+            "Table 5: Comparison between standard quantization methods. The original model has a dimensionality of 8 and 2M buckets. Note that all of the methods are without dictionary. "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>k</td><td>co</td><td colspan=\"2\">AG</td><td colspan=\"2\">Amz. f.</td><td colspan=\"2\">Amz. p.</td><td colspan=\"2\">DBP</td><td colspan=\"2\">Sogou</td><td colspan=\"2\">Yah.</td><td colspan=\"2\">Yelp f.</td><td colspan=\"2\">Yelp p.</td></tr><tr><td></td><td>full, nodict</td><td>92.1</td><td>34M</td><td>59.8</td><td>78M</td><td>94.5</td><td>83M</td><td>98.4</td><td>56M</td><td>96.3</td><td>42M</td><td>72.2</td><td>91M</td><td>63.7</td><td>48M</td><td>95.6</td><td>46M</td></tr><tr><td>8</td><td>full</td><td>92.0</td><td>9.5M</td><td>59.8</td><td>22M</td><td>94.5</td><td>24M</td><td>98.4</td><td>16M</td><td>96.3</td><td>12M</td><td>72.1</td><td>26M</td><td>63.7</td><td>14M</td><td>95.6</td><td>13M</td></tr><tr><td>4</td><td>full</td><td>92.1</td><td>5.3M</td><td>59.8</td><td>13M</td><td>94.5</td><td>13M</td><td>98.4</td><td>8.8M</td><td>96.3</td><td>6.6M</td><td>72</td><td>15M</td><td>63.6</td><td>7.5M</td><td>95.6</td><td>7.2M</td></tr><tr><td>2</td><td>full</td><td>91.9</td><td>3.2M</td><td>59.6</td><td>7.3M</td><td>94.5</td><td>7.8M</td><td>98.1</td><td>5.3M</td><td>96.3</td><td>4.0M</td><td>71.3</td><td>8.6M</td><td>63.4</td><td>4.5M</td><td>95.6</td><td>4.3M</td></tr><tr><td>8</td><td>200K</td><td>92.0</td><td>2.5M</td><td>59.7</td><td>2.5M</td><td>94.3</td><td>2.5M</td><td>98.5</td><td>2.5M</td><td>96.6</td><td>2.5M</td><td>71.8</td><td>2.5M</td><td>63.3</td><td>2.5M</td><td>95.6</td><td>2.5M</td></tr><tr><td>8</td><td>100K</td><td>91.9</td><td>1.3M</td><td>59.5</td><td>1.3M</td><td>94.3</td><td>1.3M</td><td>98.5</td><td>1.3M</td><td>96.6</td><td>1.3M</td><td>71.6</td><td>1.3M</td><td>63.4</td><td>1.3M</td><td>95.6</td><td>1.3M</td></tr><tr><td>8</td><td>50K</td><td>91.7</td><td>645K</td><td>59.7</td><td>645K</td><td>94.3</td><td>644K</td><td>98.5</td><td>645K</td><td>96.6</td><td>645K</td><td>71.5</td><td>645K</td><td>63.2</td><td>645K</td><td>95.6</td><td>644K</td></tr><tr><td>8</td><td>10K</td><td>91.3</td><td>137K</td><td>58.6</td><td>137K</td><td>93.2</td><td>137K</td><td>98.5</td><td>137K</td><td>96.5</td><td>137K</td><td>71.3</td><td>137K</td><td>63.3</td><td>137K</td><td>95.4</td><td>137K</td></tr><tr><td>4</td><td>200K</td><td>92.0</td><td>1.8M</td><td>59.7</td><td>1.8M</td><td>94.3</td><td>1.8M</td><td>98.5</td><td>1.8M</td><td>96.6</td><td>1.8M</td><td>71.7</td><td>1.8M</td><td>63.3</td><td>1.8M</td><td>95.6</td><td>1.8M</td></tr><tr><td>4</td><td>100K</td><td>91.9</td><td>889K</td><td>59.5</td><td>889K</td><td>94.4</td><td>889K</td><td>98.5</td><td>889K</td><td>96.6</td><td>889K</td><td>71.7</td><td>889K</td><td>63.4</td><td>889K</td><td>95.6</td><td>889K</td></tr><tr><td>4</td><td>50K</td><td>91.7</td><td>449K</td><td>59.6</td><td>449K</td><td>94.3</td><td>449K</td><td>98.5</td><td>450K</td><td>96.6</td><td>449K</td><td>71.4</td><td>450K</td><td>63.2</td><td>449K</td><td>95.5</td><td>449K</td></tr><tr><td>4</td><td>10K</td><td>91.5</td><td>98K</td><td>58.6</td><td>98K</td><td>93.2</td><td>98K</td><td>98.5</td><td>98K</td><td>96.5</td><td>98K</td><td>71.2</td><td>98K</td><td>63.3</td><td>98K</td><td>95.4</td><td>98K</td></tr><tr><td></td><td>200K</td><td>91.9</td><td>1.4M</td><td>59.6</td><td>1.4M</td><td>94.3</td><td>1.4M</td><td>98.4</td><td>1.4M</td><td>96.5</td><td>1.4M</td><td>71.5</td><td>1.4M</td><td>63.2</td><td>1.4M</td><td>95.5</td><td>1.4M</td></tr><tr><td></td><td>100K</td><td>91.6</td><td>693K</td><td>59.5</td><td>693K</td><td>94.3</td><td>693K</td><td>98.4</td><td>694K</td><td>96.6</td><td>693K</td><td>71.1</td><td>694K</td><td>63.2</td><td>693K</td><td>95.6</td><td>693K</td></tr><tr><td>2222</td><td>50K</td><td>91.6</td><td>352K</td><td>59.6</td><td>352K</td><td>94.3</td><td>352K</td><td>98.4</td><td>352K</td><td>96.5</td><td>352K</td><td>71.1</td><td>352K</td><td>63.2</td><td>352K</td><td>95.6</td><td>352K</td></tr><tr><td></td><td>10K</td><td>91.3</td><td>78K</td><td>58.5</td><td>78K</td><td>93.2</td><td>78K</td><td>98.4</td><td>79K</td><td>96.5</td><td>78K</td><td>70.8</td><td>78K</td><td>63.2</td><td>78K</td><td>95.3</td><td>78K</td></tr></table>",
+        "bbox": [
+            91,
+            583,
+            892,
+            832
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Comparison with different quantization and level of pruning. “co” is the cut-off parameter of the pruning. ",
+        "bbox": [
+            173,
+            847,
+            825,
+            876
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/7b583938fc3620706ee3be48ccf01b11b3457d113956cb175da280a329971ab2.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 7: Comparison between CNNs and fastText with and without quantization. The numbers for Zhang et al. (2015) are reported from Xiao & Cho (2016). Note that for the CNNs, we report the size of the model under the assumption that they use float32 storage. For fastText $( + \\mathrm { P Q } )$ we report the memory used in RAM at test time. "
+        ],
+        "table_body": "<table><tr><td>Dataset</td><td>Zhang et al. (2015)</td><td></td><td>Xiao &amp; Cho (2016)</td><td></td><td>fastText+PQ,k= d/2</td></tr><tr><td>AG</td><td>90.2</td><td>108M</td><td>91.4</td><td>80M</td><td>91.9</td></tr><tr><td>Amz. f.</td><td>59.5</td><td>10.8M</td><td>59.2</td><td>1.6M</td><td>59.6 449K</td></tr><tr><td>Amz. p.</td><td>94.5</td><td>10.8M</td><td>94.1</td><td>1.6M</td><td>449K</td></tr><tr><td>DBP</td><td>98.3</td><td>108M</td><td>98.6</td><td>1.2M</td><td>98K</td></tr><tr><td>Sogou</td><td>95.1</td><td>108M</td><td>95.2</td><td>1.6M</td><td>98K</td></tr><tr><td>Yah.</td><td>70.5</td><td>108M</td><td>71.4</td><td>80M</td><td>889K</td></tr><tr><td>Yelp f.</td><td>61.6</td><td>108M</td><td>61.8</td><td>1.4M</td><td>98K</td></tr><tr><td>Yelp p.</td><td>94.8</td><td>108M</td><td>94.5</td><td>1.2M</td><td>449K</td></tr></table>",
+        "bbox": [
+            240,
+            205,
+            753,
+            335
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/202c88cfb98e239a64d6b0def870225e77ac934cdf111c4aa5dd4309c6f7e242.jpg",
+        "table_caption": [
+            "Table 8: Comparison with and without Bloom filters. For NPQ, we set $d = 8$ and $k = 2$ . "
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Quant.</td><td>1ooom</td><td>co</td><td colspan=\"2\">AG</td><td colspan=\"2\">Amz. f.</td><td colspan=\"2\">Amz. p.</td><td colspan=\"2\">DBP</td><td colspan=\"2\">Sogou</td><td colspan=\"2\">Yah.</td><td colspan=\"2\">Yelp f.</td><td colspan=\"2\">Yelp p.</td></tr><tr><td>full,nodict</td><td></td><td></td><td>92.1</td><td>34M</td><td>59.8</td><td>78M</td><td>94.5</td><td>83M</td><td>98.4</td><td>56M</td><td>96.3</td><td>42M</td><td>72.2</td><td>91M</td><td>63.7</td><td>48M</td><td>95.6</td><td>46M</td></tr><tr><td>NPQ</td><td></td><td>200K</td><td>91.9</td><td>1.4M</td><td>59.6</td><td>1.4M</td><td>94.3</td><td>1.4M</td><td>98.4</td><td>1.4M</td><td>96.5</td><td>1.4M</td><td>71.5</td><td>1.4M</td><td>63.2</td><td>1.4M</td><td>95.5</td><td>1.4M</td></tr><tr><td>NPQ</td><td>X</td><td>200K</td><td>92.2</td><td>830K</td><td>59.3</td><td>830K</td><td>94.1</td><td>830K</td><td>98.4</td><td>830K</td><td>96.5</td><td>830K</td><td>70.7</td><td>830K</td><td>63.0</td><td>830K</td><td>95.5</td><td>830K</td></tr><tr><td>NPQ</td><td></td><td>100K</td><td>91.6</td><td>693K</td><td>59.5</td><td>693K</td><td>94.3</td><td>693K</td><td>98.4</td><td>694K</td><td>96.6</td><td>693K</td><td>71.1</td><td>694K</td><td>63.2</td><td>693K</td><td>95.6</td><td>693K</td></tr><tr><td>NPQ</td><td>X</td><td>100K</td><td>91.8</td><td>420K</td><td>59.1</td><td>420K</td><td>93.9</td><td>420K</td><td>98.4</td><td>420K</td><td>96.5</td><td>420K</td><td>70.6</td><td>420K</td><td>62.8</td><td>420K</td><td>95.3</td><td>420K</td></tr><tr><td>NPQ</td><td></td><td>50K</td><td>91.6</td><td>352K</td><td>59.6</td><td>352K</td><td>94.3</td><td>352K</td><td>98.4</td><td>352K</td><td>96.5</td><td>352K</td><td>71.1</td><td>352K</td><td>63.2</td><td>352K</td><td>95.6</td><td>352K</td></tr><tr><td>NPQ</td><td>X</td><td>50K</td><td>91.5</td><td>215K</td><td>58.8</td><td>215K</td><td>93.6</td><td>215K</td><td>98.3</td><td>215K</td><td>96.5</td><td>215K</td><td>70.1</td><td>215K</td><td>62.7</td><td>215K</td><td>95.1</td><td>215K</td></tr><tr><td>NPQ</td><td></td><td>10K</td><td>91.3</td><td>78K</td><td>58.5</td><td>78K</td><td>93.2</td><td>78K</td><td>98.4</td><td>79K</td><td>96.5</td><td>78K</td><td>70.8</td><td>78K</td><td>63.2</td><td>78K</td><td>95.3</td><td>78K</td></tr><tr><td>NPQ</td><td>X</td><td>10K</td><td>90.8</td><td>51K</td><td>56.8</td><td>51K</td><td>91.7</td><td>51K</td><td>98.1</td><td>51K</td><td>96.1</td><td>51K</td><td>68.7</td><td>51K</td><td>61.7</td><td>51K</td><td>94.5</td><td>51K</td></tr></table>",
+        "bbox": [
+            86,
+            619,
+            919,
+            791
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/aa57190841d9d805c71d4831ad3d1d6e2a74910e6d6095d3fc30a4925439025b.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 9: FlickrTag: Comparison for a large dataset of (i) different quantization methods and parameters, (ii) with or without re-training. "
+        ],
+        "table_body": "<table><tr><td>Model</td><td>k</td><td>norm</td><td>retrain</td><td>Acc.</td><td>Size</td></tr><tr><td>full</td><td></td><td></td><td></td><td>45.4</td><td>12G</td></tr><tr><td>Input</td><td>128</td><td></td><td></td><td>45.0</td><td>1.7G</td></tr><tr><td>Input</td><td>128</td><td>X</td><td></td><td>45.3</td><td>1.8G</td></tr><tr><td>Input</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>1.8G</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td></td><td>45.2</td><td>1.5G</td></tr><tr><td>Input+Output</td><td>128</td><td>X</td><td>X</td><td>45.4</td><td>1.5G</td></tr><tr><td>Input+Output, co=2M</td><td>128</td><td>X</td><td>X</td><td>45.5</td><td>305M</td></tr><tr><td>Input+Output, n co=1M</td><td>128</td><td>X</td><td>X</td><td>43.9</td><td>179M</td></tr><tr><td>Input</td><td>64</td><td></td><td></td><td>44.0</td><td>1.1G</td></tr><tr><td>Input</td><td>64</td><td>X</td><td></td><td>44.7</td><td>1.1G</td></tr><tr><td>Input</td><td>64</td><td>X</td><td></td><td>44.9</td><td>1.1G</td></tr><tr><td>Input+Output</td><td>64</td><td>X</td><td></td><td>44.6</td><td>784M</td></tr><tr><td>Input+Output</td><td>64</td><td>X</td><td>X</td><td>44.8</td><td>784M</td></tr><tr><td>Input+Output, co=2M</td><td>64</td><td>X</td><td></td><td>42.5</td><td>183M</td></tr><tr><td>Input+Output, co=1M</td><td>64</td><td>X</td><td></td><td>39.9</td><td>118M</td></tr><tr><td>Input+Output, co=2M</td><td>64</td><td>X</td><td>X</td><td>45.0</td><td>183M</td></tr><tr><td>Input+Output,co=1M</td><td>64</td><td>X</td><td>X</td><td>43.4</td><td>118M</td></tr><tr><td>Input</td><td>32</td><td></td><td></td><td>40.5</td><td>690M</td></tr><tr><td>Input</td><td>32</td><td>X</td><td></td><td>42.4</td><td>701M</td></tr><tr><td>Input</td><td>32</td><td>X</td><td>X</td><td>42.9</td><td>701M</td></tr><tr><td>Input+Output</td><td>32</td><td>X</td><td></td><td>42.3</td><td>435M</td></tr><tr><td>Input+Output</td><td>32</td><td>X</td><td>X</td><td>42.8</td><td>435M</td></tr><tr><td>Input+Output,co=2M</td><td>32</td><td>X</td><td></td><td>35.0</td><td>122M</td></tr><tr><td>Input+Output, co=1M</td><td>32</td><td>X</td><td></td><td>32.6</td><td>88M</td></tr><tr><td>Input+Output, co=2M</td><td>32</td><td>X</td><td>X</td><td>43.3</td><td>122M</td></tr><tr><td>Input+Output, co=1M</td><td>32</td><td>X</td><td>X</td><td>41.6</td><td>88M</td></tr></table>",
+        "bbox": [
+            289,
+            305,
+            707,
+            679
+        ],
+        "page_idx": 12
+    }
+]